acceptodds
Under review as a conference paper at ICLR 2027

Curvature Sign Determines Geometric Error-Propagation Regimes in Behavioral Cloning

Abstract

Behavioral cloning (BC) optimizes one-step prediction, but deployment feeds predictions back into future states, so local errors compound. We ask how rollout geometry determines the regime in which these errors accumulate without online correction. On homogeneous constant-curvature manifolds, we derive a parameter-free local recurrence separating inherited-error propagation from fresh-error injection. Under direction averaging and zero state-dependent policy amplification (), the reduced recurrence yields three regimes: a bounded squared-error floor under positive curvature, diffusive root-mean-square growth in flat space, and exponential amplification under negative curvature; a conditional orientation band extends the result beyond exact averaging. Simulations match the theory within 3% in second moment on and , and exhibits the predicted stabilization. Learned-policy evaluations on and in an articulated execution chain show that matched one-step error can coexist with different rollout errors, consistent with measured propagation response. At fixed decision/noise-injection counts, longer action chunks strengthen positive-curvature contraction and negative-curvature amplification, while the negative-curvature regime offers a geometric perspective on fragility in hyperbolic reinforcement learning. Curvature sign sets the accumulation regime, so sequential evaluation should measure propagation, not just error magnitude.

Primary area:
learning theory
Keywords:
behavioral cloning, compounding error, Riemannian geometry

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_bUVBOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how state-space curvature affects error accumulation in behavioral cloning. Using Jacobi-field analysis on constant-curvature manifolds, it separates the propagation of inherited state errors from the injection of fresh action errors. Under a local, direction-averaged oracle model with no state-dependent policy amplification, the authors derive a second-moment recurrence whose leading-order approximation exhibits three regimes: a bounded error floor under positive curvature, diffusive RMS growth in flat space, and exponential local amplification under negative curvature. The paper also provides conditional orientation bounds, studies action chunking at fixed decision/noise-injection counts, and proposes propagation diagnostics for learned policies. Validation includes controlled simulations on spheres, Euclidean and hyperbolic spaces, and SO(3), alongside learned-policy evaluations and a simulated 7-DOF Panda execution pipeline. The simulations support the recurrence within its local validity regime, while matched-one-step-error comparisons show that measured propagation response provides additional information about rollout performance.

Critical strengths
  1. Clear and quantitatively testable separation of two error mechanisms. The Jacobi-field analysis distinguishes inherited state-error propagation from fresh action-error injection, yielding different geometric coefficients for each. This provides an interpretable account of how curvature influences error accumulation under the stated local assumptions. Controlled simulations support the quantitative predictions, with reported second-moment agreement within approximately 3% on S2S^2 and R2\mathbb{R}^2. This strengthens my assessment that the proposed decomposition captures a meaningful mechanism rather than merely describing qualitative trends.

  2. Controlled evidence that propagation measurements complement one-step accuracy. The Panda evaluation extends the investigation to fully self-fed execution through inverse kinematics, joint servoing, and robot dynamics. Five matched policy pairs, repeated across three training seeds and checked on held-out calibration states, consistently associate lower measured amplification with lower rollout error. In a particularly close comparison, mean one-step errors are approximately 0.1074 and 0.1075 rad, while late-half tracking errors differ at 0.527 and 0.472 rad. These results substantiate the practical value of measuring propagation response alongside prediction accuracy and strengthen the paper’s relevance beyond its oracle model.

Critical weaknesses
  1. The learned-policy sensitivity diagnostic lacks sufficient justification and empirical support. The theoretical LL is defined intrinsically using parallel transport, whereas the experiments estimate a representation-space finite-difference sensitivity. Without a quantitative relationship between these quantities, comparisons against LcritL_{\mathrm{crit}} are difficult to interpret.

  2. The connection between the local geometric recurrence and long-horizon learned-policy behavior remains incomplete. The three regimes follow from a reduced recurrence under direction averaging, transported oracle actions, and conditionally zero-mean isotropic errors. The paper does not provide finite-horizon approximation bounds or conditions ensuring that the original process remains within the local regime. The large theory–simulation discrepancies reported in Appendix F illustrate why this matters. The learned-policy experiments provide useful supporting evidence, but the Panda comparisons hold geometry fixed and primarily establish the relevance of general propagation response. Consequently, the paper’s strongest practical results do not directly validate its curvature-specific quantitative predictions. This limits my assessment of how much the analysis establishes about error accumulation in actual behavioral cloning beyond the controlled oracle setting.

Questions for authors
  1. Clarification of the sensitivity diagnostic. How does the representation-space estimate of LL relate quantitatively to the intrinsic sensitivity defined in Section 3.1, and what makes comparisons with LcritL_{\mathrm{crit}} meaningful across representations?

  2. Clarification of the recurrence’s validity range. I understand that the three-regime result applies to the reduced local recurrence. Do the existing analysis or experiments provide a quantitative criterion, involving noise magnitude, step size, accumulated separation, and horizon, for when this recurrence accurately approximates the original process? Such a criterion would help assess the practical significance of the predicted error floor and exponential amplification, particularly given the discrepancies in Appendix F.

  3. Evidence connecting geometry to learned-policy behavior. Can the full-versus-clamped measurements in Appendix L quantify how much of the observed self-fed rollout difference is explained by geometric propagation and how much by policy response? In interpreting the learned-policy diagnostics, how are differences between calibration perturbations and actual rollout error directions and the evolving previous-action input accounted for? Analysis of the existing measurements would help establish the strength of the connection between the geometric theory and the practical results.

4 comments
AuthorsOpenReview ↗

Reply to Reviewer bUVB: Sensitivity Diagnostic, Validity Range, and Geometry in Learned Policies

We thank Reviewer bUVB for the careful review.

Weakness 1. The theoretical L is intrinsic, but the estimated one is representation-specific.

Please see Question 1.

Weakness 2. No finite-horizon bounds; Panda holds geometry fixed, so it does not test the curvature-specific predictions.

The bounds are now available; see Question 2. We agree with how you divide the evidence: Panda supports measuring propagation, and the curvature-specific evidence on learned policies is the Appendix L experiment (pp. 29-30); see Question 3. We will move it into Section 4.5 and add "Local" to the title.

Question 1. How does the representation-space L relate to the intrinsic L, and what makes comparisons with L_crit meaningful across representations?

They cannot be compared against one threshold, so we drop that comparison. We re-estimated L as Section 3.1 defines it, from decoded tangent vectors after parallel transport (95th percentile). Across the 24 configurations, the intrinsic estimate is 0.27 to 2.5 times the representation-space estimate. The sensitivity-approximation trade-off remains in all four designs: for design A, the intrinsic L grows from 0.016 to 3.51 while the one-step geodesic error falls from 0.194 to 0.115 rad.

We will redraw Figure 6(a) with intrinsic quantities, report the measured β_emp, and move β_eff and L_crit to the appendix as diagnostic proxies.

Question 2. Is there a quantitative criterion (noise, step size, separation, horizon) for when the reduced recurrence is accurate?

Yes. In "Technical Supplement, Part 1" on this page, we state and prove, under direction-averaged oracle assumptions, an identity for the oracle process itself, with no remainder dropped and no small-error restriction, in the chordal distance χ_t = χ(e_t) defined there (close to e_t for small errors):

E[χt+12∣Gt]=aχt2+bfor every t, almost surely,E[\chi_{t+1}^2 \mid \mathcal G_t] = a\chi_t^2 + b \quad \text{for every } t, \text{ almost surely,}

with explicit constants a = β_det + O(ε²) and b = σ²_noise ε² + O(ε³). The bounds below are proved in Part 2. The criterion has three parts:

  • Coefficients: the reduced recurrence uses β_det and σ²_noise ε² in place of a and b. On H² with v = 0.3 and ε = 0.01, its second moment differs from E[χ_t²] by 0.13% after 40 steps and 0.64% after 150.
  • Distance: it uses e_t² in place of χ_t². While e_t ≤ r, their relative difference is about r²/12 (2.1% at r = 0.5).
  • Exit probability: with τ_r the first time e_t exceeds r, P(τ_r ≤ T) ≤ b(a^T - 1) / ((a - 1) χ(r)²) under negative curvature. In the same example with r = 0.5, the bound is 1.3% / 4.4% / 12.2% at T = 20 / 40 / 60; 200,000 rollouts give 0 / 0 / 0.035%.

For the error floor: under positive curvature (0 ≤ a < 1), E[χ_t²] ≤ b/(1-a) at every horizon. On S² with v = 0.3 and ε = 0.01, this chordal RMS floor is 0.0475, and the simulated geodesic RMS error is 0.0478.

Appendix F's discrepancies arise where the error is far outside such a radius. In chordal distance, the theory-to-simulation ratios for the five chunk sizes of Table 3 are 1.000, 1.001, 0.999, 1.001 and 1.008 (200,000 rollouts each). We will add the theorem to Section 3.3 and a chordal row to Table 3.

Question 3. Can the full-versus-clamped measurements separate geometry from policy response? How are rollout error directions and the evolving previous-action input accounted for?

At the level of one step, yes. From the Appendix L runs at v = 0.3 (the clamp and contrast rows are new numbers from these runs; the contrast is taken per condition):

S²R²H²
β_det (oracle)0.9561.0001.046
clamp (response removed)0.9571.0001.047
full minus clamp-0.016 to -0.020-0.017 to -0.019-0.019 to -0.022

The clamped values differ across manifolds as the geometry predicts, while the full-minus-clamp contrast is similar on all three. The contrast is a signed response that includes geometry-policy interaction, not an additive share.

Over many steps the two interact, so we report the rollout outcome rather than a percentage split. The fixed-history amplification describes how an inherited error propagates locally, while a self-fed rollout also keeps adding prediction error and changes in the history input; an amplification below 1 therefore does not exclude threshold crossings. At matched one-step error, 42% / 58% / 94% of fully self-fed trajectories on S² / R² / H² cross the error threshold 1 within 150 steps.

Along the actual rollout error direction, at learner-visited states, the amplification keeps S² < R² < H² in all five error groups (0.944 to 0.990, 0.978 to 0.993, 1.015 to 1.030).

For the previous-action input, on Panda (60 held-out test trajectories), all 15 matched pairs keep the order of their self-fed late-half errors under four probes, including the actual error direction and a perturbed previous action. We will add these analyses to Appendices L and M.

I thank the authors for their detailed response and the additional analyses. Re-estimating policy sensitivity using decoded, parallel-transported tangent vectors, and dropping the unified cross-representation threshold comparison, has essentially resolved my concerns about the sensitivity diagnostic. Measuring along the true error directions on states actually visited by the learned policy also makes the connection between local propagation response and self-fed execution behavior clearer.

Regarding the validity range of the recurrence, the exact chord-distance recurrence and the exit-probability bound provided by the authors directly address my main concern. I will examine the theorem statements and proofs in the technical supplement more closely. I would like the revision to distinguish explicitly between two kinds of results: the global conclusions established for chord distance under the direction-averaged oracle model, and the conclusions established for geodesic error at a given radius and horizon. Concretely, do the present results guarantee that trajectories remain within the region where the local approximation is valid with high probability, or do they also control the contribution of trajectories that exit this region to the unconditional second moment of geodesic error, E ⁣[et2]\mathbb{E}\!\left[e_t^{2}\right]?

As for the Panda experiments, the fact that the 15 matched policy pairs preserve identical rankings across different probing methods further supports the empirical value of the propagation diagnostic. I would also ask the authors to explain how the previous-action perturbation is constructed, in particular, whether its magnitude and direction reflect the action-history discrepancies that actually arise between the learner and the expert in self-fed execution. This response has substantially alleviated my earlier concerns; the remaining question mainly concerns the scope within which these quantitative conclusions apply to more general learned policies.

AuthorsOpenReview ↗

Follow-up Reply to Reviewer bUVB: Chordal vs. Geodesic Guarantees and the Previous-Action Probe

Thank you for the careful follow-up. We will make the following distinctions and the probe construction explicit in the revision.

1. Global moments and local validity. Under the direction-averaged oracle assumptions of "Technical Supplement, Part 1", the recurrence for E[χt2]E[\chi_t^2] is exact, with no small-error restriction. On HdH^d, the exit-probability bound of Part 2 guarantees that trajectories stay within radius rr through horizon TT with the stated probability. By itself, it does not control what exiting trajectories contribute to the unconditional geodesic second moment.

We also obtain unconditional bounds on the geodesic second moment that include trajectories leaving the local region. Two of them are already in Part 2: on SdS^d, E[et2]≤(π2/4)b/(1−a)E[e_t^2]\le(\pi^2/4)b/(1-a) when 0≤a<10\le a\lt1; on Rd\mathbb R^d, E[et2]=tε2E[e_t^2]=t\varepsilon^2. The bound on HdH^d is new. It uses only the identity of Part 1 and Jensen's inequality, with no further assumption. Let yt=E[χt2]y_t=E[\chi_t^2] be the exact chordal second moment of Part 2. Since χ(e)2=2(cosh⁡e−1)\chi(e)^2=2(\cosh e-1) is an increasing convex function of e2e^2, Jensen's inequality gives 2(cosh⁡E[et2]−1)≤E[χt2]=yt2(\cosh\sqrt{E[e_t^2]}-1)\le E[\chi_t^2]=y_t. Solving for E[et2]E[e_t^2] gives

E[et2]≤(2sinh⁡−1yt2)2≤yt.E[e_t^2]\le\Big(2\sinh^{-1}\frac{\sqrt{y_t}}{2}\Big)^2\le y_t .

These bounds control the total second moment; they do not establish that the contribution from exiting trajectories is negligible. On H2H^2 with v=0.3v=0.3 and ε=0.01\varepsilon=0.01 (200,000 rollouts), the resulting bound on the geodesic RMS error is 1.00, 1.01, 1.05, 1.10, 1.08, 1.06 times the simulated RMS error at t=40,100,150,200,300,400t=40,100,150,200,300,400; the corresponding RMS prediction from the reduced recurrence is 1.00, 1.02, 1.12, 1.57, 6.06, 35.6 times the simulated RMS error.

We do not claim a matching lower bound on HdH^d. The small-error expansion connects the exact recurrence to the paper's reduced recurrence, whose accuracy is assessed numerically. The unconditional upper bound above remains valid beyond the local regime. In the revision, we will add this bound to the results of Part 2 and present three kinds of statement separately: the global chordal identity, the unconditional geodesic bounds, and the local results (the distance comparison within a radius and the probability of staying there).

2. Previous-action perturbation. Its direction is taken from the actual learner-expert history discrepancy in self-fed execution. Let hth_t be the learner's previous output, which is fed back as an input, and ht∗h_t^\ast the expert's previous action encoded in the same policy-specific input representation. At probed states with tracking error et≥0.02e_t\ge0.02 rad, we perturb the end-effector orientation toward the expert by δ=0.01\delta=0.01 rad and, at the same time, replace hth_t by

ht′=ht+δet(ht∗−ht).h'_t=h_t+\frac{\delta}{e_t}(h_t^\ast-h_t).

The history input therefore moves along the observed discrepancy, by the same fraction δ/et\delta/e_t as the state. This makes it a local probe, not a full-size replay of the discrepancy. The perturbed orientation is realized through IK, and the probe uses the same servo/MuJoCo execution chain as the original measurement.

This joint probe preserves the reported ordering in all 15 matched policy pairs. So does a full-amplitude version, in which the policy is executed from the expert's state with the expert's previous action as input. Both probes move the state and the history together, so they test whether the ordering holds, not how much the history input contributes by itself. We will describe this construction in the revision and present the consistent ranking as an empirical robustness check of the diagnostic on the tested policies.

3. Scope for learned policies. The quantitative theorems are for the direction-averaged oracle; for learned policies our claim is diagnostic. Within the modeled geodesic dynamics, the geometric step map conditional on the executed action applies regardless of how that action is selected; Appendix L's clamp test agrees with Remark 7 to within 2×10−62\times10^{-6}. The propagation response βemp\beta_{\rm emp} can also be measured through the full execution chain. The policy response is measured empirically rather than predicted by the oracle theory. The association between the measured propagation response and rollout-error rankings is empirical: the three model spaces of Appendix L, SO(3), and the Panda chain. The revised abstract, Section 4.5 and limitations will state this scope explicitly.

Thank you for the detailed follow-up and the additional analyses. The distinction among the global chordal identity, the unconditional geodesic upper bounds, and the local validity results clarifies the scope of the theoretical claims. The explanation of the previous-action perturbation, together with the consistent rankings under both local and full-amplitude probes, also addresses my question about how the diagnostic accounts for action-history discrepancies.

Overall, the rebuttal has largely addressed my concerns. I appreciate the explicit separation between the theoretical guarantees for the direction-averaged oracle model and the empirical evidence for learned policies, and I encourage the authors to incorporate these clarifications and the additional results into the revised manuscript. These responses strengthen my confidence in the paper, and I maintain my original rating.

Reviewer_JtoxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how state-space curvature affects the propagation of local prediction errors in behavioural cloning. On a homogeneous constant-curvature Riemannian manifold with geodesic state updates, fixed action norm, direction averaging, and an L=0 oracle learner, the authors use Jacobi fields to derive separate one-step factors for inherited state separation and newly injected action error. Dropping higher-order terms gives a scalar second-moment recurrence whose inherited-error coefficient is below, equal to, or above one under positive, zero, or negative curvature, respectively. A conditional orientation band partially relaxes exact direction averaging.

The paper validates the local recurrence on S2, R2, H2, and SO(3), studies a fixed-decision-count action-chunking protocol and a negative-curvature surrogate, and evaluates propagation diagnostics for learned policies on SO(3), on the three model spaces, and through a simulated 7-DOF Panda execution chain. The empirical message is that closely matched one-step error can coexist with different rollout error, and that directly measured local state amplification can provide information beyond one-step accuracy.

Critical strengths
  1. Clean mechanism-level geometric decomposition. The distinction between inherited separation and fresh action error is mathematically useful: the two channels correspond to different Jacobi-field initial conditions and therefore produce different curvature-dependent factors. The assumptions are stated relatively clearly, and the local L=0 derivation appears internally consistent. This is more informative than collapsing all compounding effects into a single undifferentiated stability coefficient.

  2. Broad and unusually transparent validation. The paper checks the recurrence on canonical constant-curvature spaces, SO(3), a variable-curvature torus diagnostic, learned policies, and an articulated robot execution chain. Importantly, it also reports where the reduced recurrence ceases to be quantitatively reliable under strong negative curvature rather than hiding this failure. The matched-error Panda comparisons are particularly useful: they provide evidence that measured propagation response can resolve rollout differences after one-step error has been closely controlled.

Critical weaknesses
  1. The title-level “three regimes” claim is established only for a reduced local recurrence, not for the actual long-horizon rollout process. Theorem 4 is a one-step expansion with an O(et3+etϵ2+ϵ3)O(e_t^3 + e_t \epsilon^2 + \epsilon^3) remainder, while Corollary 5 obtains the bounded/diffusive/exponential trichotomy by dropping that remainder. There is no theorem controlling the accumulated remainder or showing that the process remains in the local neighbourhood for the required horizon. In fact, with bounded action lengths, the geodesic tracking distance itself is at most linear in the horizon by the triangle inequality, so “exponential” cannot describe its asymptotic growth. It is instead a transient linearised amplification regime before the process leaves the local neighbourhood. This distinction is consequential because several statements in the title, abstract, and conclusion read as claims about the accumulation regime of the actual rollout. The appendix confirms the issue quantitatively: for strong negative curvature or large effective steps, the reduced recurrence can overpredict the simulated RMS error by many orders of magnitude.

  2. The bridge from the formal oracle to deployed behavioural cloning remains incomplete. In the L=0L=0 model, the learner receives the expert action at the expert state at every time step and parallel-transports it to the learner state before adding fresh isotropic noise. This is a useful synchronous oracle coupling, but it is not the usual deployed learner action $ \hat{\pi} ( \hat{s}_t ) $.

For learned policies, βeff\beta_{eff} and LcritL_{crit} are explicitly heuristic, and LL is estimated in each architecture’s own input/output representation. These representation-specific quantities are not directly comparable to the intrinsic tangent-space sensitivity in the theory, yet Figure 6(a) places different representations against common LcritL_{crit} thresholds and the text compares their LL values. The direct intrinsic state-amplification measurement βemp\beta_{emp} is substantially better grounded than this LL-based comparison. The paper’s strongest learned-policy evidence therefore supports the general value of measuring propagation, but does not yet establish that curvature sign alone determines ordinary BC error accumulation.

Questions for authors
  1. Can the authors provide a finite-horizon or stopping-time version of the regime result for the actual stochastic process, rather than only for the reduced recurrence? For example, for κ<0\kappa < 0, can they bound the horizon over which ete_t remains local and the accumulated remainder remains smaller than the leading βdett\beta_{det}^t term as a function of epsilon, βdet\beta_{det}, and a prescribed neighbourhood radius? If such a result is not available, would the authors revise the title and the main claims to say explicitly “local/transient geometric amplification regimes” rather than long-horizon accumulation regimes?

  2. How should the representation-space finite-difference estimate LL used in Experiment 5 be related quantitatively to the intrinsic manifold-level LL in Section 3.1? In particular, what norm-equivalence or decoder-Jacobian constants justify plotting the 6D and geodesic representations against the same LcritL_{crit} thresholds? Would the main learned-policy conclusion remain unchanged if Figure 6(a) were reformulated entirely in common intrinsic quantities, such as decoded geodesic one-step error and βemp\beta_{emp}?

  3. The chunking result holds under a fixed decision/noise-injection count, so increasing HH also increases the total primitive motion or path length represented by the rollout. What happens under a fixed primitive-step horizon or fixed total physical path length, where the number of decisions decreases with HH? This comparison seems necessary before interpreting the result as a claim about practical action chunking rather than about longer geodesic steps.

2 comments
AuthorsOpenReview ↗

Reply to Reviewer Jtox: Stopping-Time Bound, Intrinsic Sensitivity, and Fixed-Horizon Chunking

We thank Reviewer Jtox for the detailed review.

Weakness 1. The three regimes are shown only for a reduced recurrence, and geodesic distance cannot grow exponentially forever.

We agree; please see Question 1.

Weakness 2. The oracle is not the deployed policy, and curvature sign alone is not shown to determine error accumulation in ordinary BC.

The oracle switches off the policy's response to state shifts, and Appendix L (pp. 29-30) measures that response on learned policies. Its "clamp" test removes only that response: at a perturbed state, the policy executes the parallel-transported action it output at the unperturbed state. With the response removed, the one-step amplification matches the geometric prediction (0.957 / 1.000 / 1.047 on S² / R² / H² at v = 0.3); restoring the response lowers it by 0.016 to 0.022 (new numbers from the Appendix L runs, per condition).

We agree with how you divide the evidence. The curvature-specific evidence is the matched-error, fully self-fed experiment of Appendix L: 42% / 58% / 94% of trajectories on S² / R² / H² cross the error threshold 1 within 150 steps, in this order in all three seeds. An amplification below 1 does not exclude crossings: a self-fed rollout also keeps adding prediction error and changes in the history input. Our conclusions concern the geometric propagation channel, not BC in general. We will move these results into Section 4.5.

Question 1. Is there a stopping-time result for the actual process: how long does e_t stay local, in terms of ε, β_det and a radius? If not, will you revise the title?

We accept the triangle-inequality argument: e_t cannot grow exponentially over an unbounded horizon. The paper's exponential regime describes local propagation. We will change the title to "Curvature Sign Determines Local Geometric Error-Propagation Regimes in Behavioral Cloning". The abstract's last sentence will begin "Under the stated oracle assumptions, curvature sign governs local geometric propagation", and the conclusion will be revised to match.

We can now give the stopping-time result for the actual oracle process, with no remainder dropped. In "Technical Supplement, Part 1" on this page, we state and prove, under direction-averaged oracle assumptions, an identity with no small-error restriction, in the chordal distance χ_t = χ(e_t) (on H², χ(e) = 2 sinh(e/2), close to e for small errors):

E[χt+12∣Gt]=aχt2+bfor every t, almost surely,E[\chi_{t+1}^2 \mid \mathcal G_t] = a\chi_t^2 + b \quad \text{for every } t, \text{ almost surely,}

with constants a = β_det + O(ε²) and b = σ²_noise ε² + O(ε³).

Let τ_r be the first time e_t exceeds a radius r. For negative curvature,

P(τ_r ≤ T) ≤ b(a^T - 1) / ((a - 1) χ(r)²).

It is proved in Part 2, with the bounds for zero and positive curvature. The bound stays below a target p as long as T ≤ ln(1 + p(a-1)χ(r)²/b) / ln a, a horizon in ε, β_det and r. On H² with v = 0.3, ε = 0.01 and r = 0.5, the bound is 4.4% at T = 40 and 12.2% at T = 60; 200,000 rollouts give 0 and 0.035%.

The theorem and these bounds will go in Section 3.3.

Question 2. How does the representation-space L relate to the intrinsic L, and does the main learned-policy conclusion survive a fully intrinsic Figure 6(a)?

We agree, and the conclusion survives. We re-estimated L as Section 3.1 defines it, from decoded tangent vectors after parallel transport (95th percentile). Across the 24 configurations, the intrinsic estimate is 0.27 to 2.5 times the representation-space estimate, so the representation-space estimates cannot be compared against one threshold. We will redraw Figure 6(a) with the intrinsic L and the geodesic one-step error, drop L_crit as a threshold, and report the measured β_emp. The sensitivity-approximation trade-off remains in all four designs: as the constraint is relaxed, L increases and the one-step error decreases.

Figure 6(b)'s observed rollout differences at matched one-step error do not involve L and remain unchanged. Using the paper's 95th-percentile summary, intrinsic sensitivity retains B > A (22.2 vs 3.5) and C > D (2.45 vs 2.01); C/D have similar means and medians. We will qualify the sensitivity interpretation accordingly.

Question 3. The chunking result fixes the number of decisions. What happens at a fixed number of primitive steps?

We fix 120 primitive steps of length 0.1, so chunk length H means 120/H decisions. The table gives the terminal E[e²]/ε² from the reduced recurrence, with the same per-decision noise ε for every H (simulation agrees within 5%).

H12358
S²9035197.93.5
R²12060402415
H²164116112155412

The three curvatures still behave differently. On R² the error is 120/H: longer chunks only mean fewer noise injections. On S² it falls faster than 1/H. On H² it is smallest at H = 3 and grows again for longer chunks. We will add this to Corollary 8 and Section 4.2.

Thank you for the detailed response and additional analyses. The exact chordal recurrence, finite-horizon bounds, intrinsic sensitivity re-estimation, and fixed-primitive-horizon chunking analysis address several of my technical questions.

However, my main reservation concerns the scope and significance of the contribution rather than only the missing bounds. The quantitative theory remains specific to a direction-averaged L=0L=0 oracle that executes the transported expert action, thereby removing the state-dependent policy response central to deployed behavioural cloning. The revised title and claims appropriately narrow the result to local geometric propagation, but this also makes clear that the formal contribution does not establish error-accumulation regimes for BC policies in general.

The learned-policy experiments provide useful empirical evidence that measured propagation response can complement one-step error. Nevertheless, these results remain diagnostic rather than theoretically explained by the oracle result: the Panda experiment holds geometry fixed, while the cross-manifold experiments do not isolate curvature sufficiently to support a general curvature-determines-BC-regime conclusion.

I therefore view the rebuttal as substantially improving the accuracy and presentation of the paper without overcoming the central limitation in its theoretical–empirical scope. I maintain my rating of 2 (weak rejection).

Reviewer_qnxfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how state tracking error in behavior cloning is affected by the geometric structure of the state space. The authors construct a local analysis model on manifolds of constant sectional curvature: the learner executes the expert action after parallel transport and superimposes a zero-mean, isotropic small perturbation, so as to exclude additional policy responses to state deviations.Using two sets of Jacobi fields under different initial conditions, the paper derives the propagation coefficient of existing state error and the injection coefficient of newly added action error, respectively, and obtains a local squared-error recursion. Under direction averaging, L=0, and neglect of higher-order remainder terms, the simplified recursion corresponds to three regimes: a finite error plateau under positive curvature, linear growth of squared error under zero curvature, and exponential amplification under negative curvature. The paper further gives upper and lower propagation bounds under conditional directional statistical constraints, and discusses action chunking and policy sensitivity.The experiments include controlled simulations in constant-curvature spaces,SO(3) learned policies, fixed-action interventions, fully self-feedback execution, and simulations of a 7-DoF Panda robotic arm. The results support the local geometric propagation mechanism and show that, when one-step errors are similar, the empirical propagation response can provide additional information about long-term performance.

Critical strengths
  1. The error decomposition is clear and provides a testable geometric mechanism. The paper distinguishes two perturbations: "different starting points with parallel initial actions" and "same starting point with different actions," and obtains responses involving cκ(v) and sκ(v), respectively. It then treats error propagation and error injection separately, clearly revealing the different roles of longitudinal and transverse components. In addition, the fixed-action intervention in Appendix L removes the differential policy response while preserving the actual actions of the learned policy, further supporting the geometric propagation formula under the angular condition.

  2. The experiments matching one-step error are highly convincing. The Panda experiment uses independent matching sets and held-out calibration sets for selection, validates policy pairs, and executes fully self-feedback rollouts under three training seeds. In the results, all 15 matched pairs exhibit consistent rankings of empirical propagation coefficients and long-term tracking error, supporting the evaluation view that propagation response must be considered beyond one-step error.The authors also report that one-step error can still explain the main performance trends in the full policy sweep, and they do not describe it as an invalid metric.Significance for the overall evaluation: Compared with merely reporting overall correlations, the matching experiment more directly supports the additional evaluation value of propagation response.

Critical weaknesses
  1. The scope of the formal theory in explaining actual long-term error patterns of BC remains limited. The oracle setting excludes a key feedback mechanism: the learner does not independently select a reference action based on the shifted state, but instead receives the current expert action after parallel transport. This can isolate geometric factors, but it also removes the key BC error accumulation mechanism in which "state shift induces changes in policy output." In addition, scalar propagation relies on direction averaging; for a given expert policy, propagation actually depends on the angle between the error and the action. The fixed-geodesic experiment in Appendix C shows that under positive curvature, the saturation predicted by the direction-averaged model does not always appear. Meanwhile, the long-term conclusions come from the recursion after deleting remainder terms; the bounded, diffusive, and exponential regimes are conclusions obtained after neglecting higher-order remainder terms. The paper does not prove that the original process remains within the local valid region over the corresponding time scale, nor does it provide control of how approximation error accumulates over time.

  2. The learned-policy experiments have not yet sufficiently established the geometric interpretation and predictive power of the propagation metric. The sensitivity proxy is not a rigorous bound; the paper uses β_eff = β_det(1 + Lv)² and the corresponding nominal threshold as diagnostic tools. There are clear discrepancies between the proxy and measurements: in Appendix I, Table 5, the proxy values exceed 1 in multiple configurations, while all measured propagation coefficients remain below 1; in some configurations, when the estimated sensitivity increases, the measured propagation coefficient instead decreases. The paper also does not clearly argue under what conditions this local, fixed-history average propagation response is sufficient to predict full closed-loop behavior; there remains a clear gap between the theoretical analysis and the actual policy evaluation.

Questions for authors
  1. Theorems 4 through Corollary 5 delete remainder terms that depend on the current error. Can an explicit finite-time error bound, a guarantee before exiting the local region, or a relationship between noise scale and valid time range be provided? If not, would the authors agree to explicitly restrict the three long-term regimes to properties of the simplified local model?

  2. The current assumption of uniform directions is too strong. Can the existing experiments be used to group by current error magnitude and task stage, and to examine the relationship between conditional angle statistics and propagation coefficients? Does the marginal averaging in Appendix L mask certain state regions that clearly deviate from the direction-averaged prediction?

  3. In Table 5, βeff\beta_{eff} and βemp\beta_{emp} are inconsistent in the critical-value judgments and in some trends. How do the authors determine whether the nominal threshold L_crit has operational meaning? If action sensitivity under a unified geometric metric were used instead of representation-space sensitivity, would the results improve?

1 comment
AuthorsOpenReview ↗

Reply to Reviewer qnxf: Finite-Time Guarantees, Direction Statistics, and the Sensitivity Proxy

We thank Reviewer qnxf for the detailed review.

Weakness 1. The oracle removes the policy's reaction to state shifts, relies on direction averaging, and drops the remainder.

The oracle isolates the geometric channel, and the clamp test of Appendix L measures the policy's response on learned policies. With the response removed, the one-step amplification at v = 0.3 is 0.957 / 1.000 / 1.047 on S² / R² / H², matching the geometric prediction; restoring the response lowers it by 0.016 to 0.022 (new numbers from the Appendix L runs, per condition). For direction averaging and the remainder, see Questions 2 and 1.

Weakness 2. β_eff disagrees with Table 5; when does a fixed-history local response predict closed-loop behavior?

For β_eff, see Question 3. In our tests the measured response β_emp gives the order of rollout errors, not their size. On Panda, all 15 matched pairs keep the order of their self-fed late-half errors under four probes, including the actual error direction and a perturbed previous action.

It does not give the size: the fixed-history amplification describes how an inherited error propagates locally, while a self-fed rollout also keeps adding prediction error and changes in the history input. So an amplification below 1 does not exclude crossings. At matched one-step error, 42% / 58% / 94% of self-fed trajectories on S² / R² / H² cross the error threshold 1 within 150 steps, in the order of the measured amplification.

Question 1. Is there a finite-time bound, a guarantee before leaving the local region, or a noise-time relation? If not, will you restrict the regimes to the local model?

We can now give all three, for the oracle process itself with no remainder dropped. In "Technical Supplement, Part 1" on this page, we state and prove, under direction-averaged oracle assumptions, an identity with no small-error restriction, in the chordal distance χ_t = χ(e_t) defined there (close to e_t for small errors):

E[χt+12∣Gt]=aχt2+bfor every t, almost surely,E[\chi_{t+1}^2 \mid \mathcal G_t] = a\chi_t^2 + b \quad \text{for every } t, \text{ almost surely,}

with constants a = β_det + O(ε²) and b = σ²_noise ε² + O(ε³). The bounds below are proved in Part 2. Let τ_r be the first time e_t exceeds a radius r.

  • Positive curvature (0 ≤ a < 1): E[e_t²] ≤ (π²/4) b/(1-a) for every t.
  • Zero curvature: E[e_t²] = tε².
  • Negative curvature: P(τ_r ≤ T) ≤ b(a^T - 1) / ((a - 1) χ(r)²).

The last bound stays below a target p as long as T ≤ ln(1 + p(a-1)χ(r)²/b) / ln a. On H² with v = 0.3, r = 0.5 and p = 5%, T = 42 for ε = 0.01 and T = 140 for ε = 0.001. The bound is conservative (12.2% at T = 60, against 0.035% in 200,000 rollouts); T = 42 and 140 are sufficient horizons for the stated 5% target.

We accept the restriction you propose: the title will say "Local Geometric Error-Propagation Regimes", with the abstract and conclusion revised to match.

Question 2. Do statistics grouped by error size and task stage show regions hidden by the marginal average?

Yes: on S², the error aligns with the action as it grows. This changes how far the one-step amplification is from 1, not which side of 1 it lies on. At learner-visited states, we measured cos²θ between error and action and the amplification along the error direction, in five error groups. On S², cos²θ rises from 0.56 in the smallest group to 0.91 in the largest. Across the groups the amplification is 0.944 to 0.990 on S², 0.978 to 0.993 on R² and 1.015 to 1.030 on H². The order S² < R² < H² holds in every error group and in each of four task stages.

Appendix C's straight-geodesic case follows the same angle dependence: the error component along the action is not contracted. The predicted total RMS error over the final 30 steps is 0.084 on S², against 0.086 measured.

Question 3. β_eff vs. β_emp in Table 5: is L_crit meaningful, and would a unified metric help?

We agree that L_crit cannot serve as a common threshold across representations for these policies; we will drop it and report the measured β_emp. β_eff = β_det(1 + Lv)² departs from β_emp for two reasons: the realized action is shorter than the nominal step, and the policy's response has a sign. With the response removed, the amplification is 0.985 to 1.000 in all 12 configurations, against a nominal β_det of 0.847 for H = 5. The full-minus-clamp contrast is negative, -0.019 to -0.131, and its magnitude increases with the constraint scale in every design, so β_emp falls while β_eff rises. Since the clamped value is within 0.015 of 1, the values below 1 in Table 5 come mainly from the policy's response; curvature-specific evidence is in Appendix L.

A unified metric makes the designs comparable, and we adopt it; the sensitivity-approximation trade-off remains in all four designs (design A: L from 0.016 to 3.51, error from 0.194 to 0.115 rad). We will redraw Figure 6(a), add clamp columns to Table 5, and move β_eff and L_crit to the appendix, revising Sections 4.3 and 4.5.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Technical Supplement, Part 1: Exact Chordal Recurrence (Statement and Proof)

For the paper's direction-averaged L=0L=0 oracle, the squared chordal error obeys an exact affine recurrence. This comment states and proves it; Part 2 derives the finite-horizon bounds.

Setting. Let M=Sd,RdM=S^d,\mathbb R^d, or HdH^d, with κ=1,0,−1\kappa=1,0,-1, respectively (other curvature values follow by rescaling lengths), d≥2d\ge2, and fixed step length v>0v\gt0. Use the paper's cκ(x)=(cos⁡x,1,cosh⁡x)c_\kappa(x)=(\cos x,1,\cosh x) and sκ(x)=(sin⁡x/x,1,sinh⁡x/x)s_\kappa(x)=(\sin x/x,1,\sinh x/x), respectively, with sκ(0)=1s_\kappa(0)=1. Write Sκ(x)=xsκ(x)S_\kappa(x)=xs_\kappa(x) and

χ(e)=2Sκ(e/2),χt=χ(et).\chi(e)=2S_\kappa(e/2),\qquad \chi_t=\chi(e_t).

Thus χ(e)2=2[1−cκ(e)]/κ\chi(e)^2=2[1-c_\kappa(e)]/\kappa for κ≠0\kappa\ne0, and χ(e)=e\chi(e)=e for κ=0\kappa=0.

At step tt, write p=st∗p=s_t^\ast, q=s^tq=\hat s_t, and et=d(p,q)e_t=d(p,q). Let ut∈TpMu_t\in T_pM denote the expert action sampled under (A1) below (we write utu_t because aa denotes a coefficient), and let Γ=Γp→q\Gamma=\Gamma_{p\to q} denote parallel transport along the minimizing geodesic. The updates are

p+=exp⁡p(ut),q+=exp⁡q(Γut+ξt),∥ut∥=v.p^+=\exp_p(u_t),\qquad q^+=\exp_q(\Gamma u_t+\xi_t),\qquad \lVert u_t\rVert=v.

All norms and angles are Riemannian. Let Gt\mathcal G_t be the history before utu_t is drawn, including the current states p,qp,q (it contains the Gt\mathcal G_t of Proposition 6). Assume:

(A1) Given Gt\mathcal G_t, ut/vu_t/v is uniform on the unit sphere of TpMT_pM.

(A2) For noise scale ε>0\varepsilon\gt0, given Gt\mathcal G_t and utu_t, ξt=εzt\xi_t=\varepsilon z_t, where ztz_t has a fixed rotationally invariant law in an orthonormal frame of TqMT_qM, with E∥zt∥2=1E\lVert z_t\rVert^2=1 and bounded support. We use the rotationally invariant case of the noise model in Section 3.1, as in the paper's simulations.

On SdS^d, take 0<v<π0\lt v\lt\pi and ε\varepsilon small enough that v+∥ξt∥<πv+\lVert\xi_t\rVert\lt\pi. Start from e0=0e_0=0. Antipodal events have probability zero: given the action, q+=−p+q^+=-p^+ requires one specific nonzero noise vector, and for d≥2d\ge2 the noise direction is continuously distributed. At coincident points, transport is the identity and the separation direction may be chosen arbitrarily before drawing the action.

Coefficients and statement. For κ=±1\kappa=\pm1, set uˉ=Γut/v\bar u=\Gamma u_t/v, ξ∥=⟨ξt,uˉ⟩\xi_\parallel=\langle\xi_t,\bar u\rangle (Riemannian inner product), and R=∥vuˉ+ξt∥R=\lVert v\bar u+\xi_t\rVert. At fixed q,utq,u_t, define the noise expectations

U=E[cκ(R)],V=E[Sκ(R)R(v+ξ∥)],U=E[c_\kappa(R)],\qquad V=E\left[\frac{S_\kappa(R)}R(v+\xi_\parallel)\right],

using Sκ(R)/R=1S_\kappa(R)/R=1 at R=0R=0. By (A2), these depend only on κ,d,v,ε\kappa,d,v,\varepsilon and the noise law. Then, for every tt, almost surely and with no small-error restriction,

E[χt+12∣Gt]=aχt2+b,E[\chi_{t+1}^2\mid\mathcal G_t]=a\chi_t^2+b,

where the scalars a,ba,b are

b=\frac2\kappa[1-c_\kappa(v)U-\kappa S_\kappa(v)V].$$ For $\kappa=0$, $a=1$ and $b=\varepsilon^2$. **Proof.** The flat case follows from $q^+-p^+=(q-p)+\xi_t$ and conditional zero-mean noise. For $\kappa=\pm1$, use the Euclidean/Minkowski ambient bilinear form $B_\kappa$, normalized so that $$B_\kappa(x,y)=c_\kappa(d(x,y)),\qquad B_\kappa(X,Y)=\kappa\langle X,Y\rangle$$ for points $x,y\in M$ and tangent vectors $X,Y$ at the same point. Suppress $t$, put $e=e_t$, and abbreviate $c_x=c_\kappa(x)$, $S_x=S_\kappa(x)$. Choose a unit tangent $f_1$ at $p$ toward $q$. Let $\theta$ be the angle between $u_t$ and $f_1$, and choose an orthogonal unit tangent $f_2$ so that $u_t/v=\cos\theta f_1+\sin\theta f_2$ (any orthogonal $f_2$ if $\sin\theta=0$). Then $$q=c_ep+S_ef_1,\qquad p^+=c_vp+S_v(\cos\theta f_1+\sin\theta f_2),$$ $$\bar u=\cos\theta(-\kappa S_ep+c_ef_1)+\sin\theta f_2.$$ Rotational symmetry around $\bar u$ cancels the transverse noise contribution in $q^+=c_\kappa(R)q+[S_\kappa(R)/R](v\bar u+\xi_t)$, giving $$E[q^+\mid\mathcal G_t,u_t]=Uq+V\bar u.$$ Direct expansion yields $$B_\kappa(p^+,q)=c_vc_e+\kappa S_vS_e\cos\theta,$$ $$B_\kappa(p^+,\bar u)=-\kappa c_vS_e\cos\theta+ \kappa S_v(c_e\cos^2\theta+\sin^2\theta).$$ By (A1), $E[\cos\theta\mid\mathcal G_t]=0$ and $E[\cos^2\theta\mid\mathcal G_t]=1/d$. Since $c_\kappa(e_{t+1})=B_\kappa(p^+,q^+)$ is linear in $q^+$, $$E[c_\kappa(e_{t+1})\mid\mathcal G_t] =\left(c_vU+\frac{\kappa}{d}S_vV\right)c_e +\kappa\frac{d-1}{d}S_vV.$$ Substituting $\chi(e)^2=2[1-c_\kappa(e)]/\kappa$ proves the recurrence. For $e_t=0$ the same computation applies with any choice of $f_1$. **Relation to the paper.** At zero noise, $U=c_\kappa(v)$ and $V=S_\kappa(v)$, so $a=\beta_{\rm det}=1/d+(d-1)c_\kappa(v)^2/d$. Smoothness and zero-mean noise give $a=\beta_{\rm det}+O(\varepsilon^2)$. Also, $b$ is the expected squared chordal displacement caused by noise at a coincident start. Lemma 2 and $\chi(e)^2=e^2+O(e^4)$ therefore give $$b=\sigma_{\rm noise}^2\varepsilon^2+O(\varepsilon^3),\qquad \sigma_{\rm noise}^2=\frac1d+\frac{d-1}{d}s_\kappa(v)^2.$$ Under these assumptions, the small-error expansion recovers Theorem 4.
AuthorsOpenReview ↗

Technical Supplement, Part 2: Finite-Horizon Bounds and Numerical Checks

Use the notation and assumptions of Part 1, with e0=0e_0=0.

Exact second moment. Define yt=E[χt2]y_t=E[\chi_t^2]. Taking expectations gives yt+1=ayt+by_{t+1}=ay_t+b, y0=0y_0=0, hence

yt=b(at−1)a−1(a≠1),yt=bt(a=1).y_t=\frac{b(a^t-1)}{a-1}\quad(a\ne1),\qquad y_t=bt\quad(a=1).

Positive curvature: a uniform-in-time mean-square bound. On SdS^d, fixed d≥2d\ge2 and 0<v<π0\lt v\lt\pi give βdet<1\beta_{\rm det}\lt1, so 0<a<10\lt a\lt1 for sufficiently small ε\varepsilon. Whenever 0≤a<10\le a\lt1,

yt≤b1−a,yt⟶b1−a.y_t\le\frac b{1-a},\qquad y_t\longrightarrow\frac b{1-a}.

For every e∈[0,π]e\in[0,\pi], (2/π)e≤χ(e)=2sin⁡(e/2)≤e(2/\pi)e\le\chi(e)=2\sin(e/2)\le e. Therefore

E[et2]≤π24b1−afor every t.E[e_t^2]\le\frac{\pi^2}{4}\frac b{1-a}\quad\text{for every }t.

Since b=O(ε2)b=O(\varepsilon^2) and 1−a→1−βdet>01-a\to1-\beta_{\rm det}\gt0 as ε→0\varepsilon\to0 at fixed vv, this is a uniform O(ε2)O(\varepsilon^2) bound. The exact limiting floor b/(1−a)b/(1-a) concerns the chordal second moment; the displayed inequality is an upper bound for the geodesic second moment.

Flat space. On Rd\mathbb R^d, χ(e)=e\chi(e)=e, a=1a=1, and b=ε2b=\varepsilon^2, so E[et2]=tε2E[e_t^2]=t\varepsilon^2 exactly.

Exit probability on Rd\mathbb R^d and HdH^d. For r>0r\gt0, let τr\tau_r be the first time t≥0t\ge0 at which et>re_t\gt r (τr=∞\tau_r=\infty if this never happens). Since b≥0b\ge0, if a≥1a\ge1 then Part 1 gives

E[χt+12∣Gt]=aχt2+b≥χt2.E[\chi_{t+1}^2\mid\mathcal G_t]=a\chi_t^2+b\ge\chi_t^2.

Thus χt2\chi_t^2 is a nonnegative submartingale. Bounded noise ensures integrability at every finite horizon. Since χ\chi is increasing, Doob's maximal inequality gives, for every integer T≥0T\ge0,

P(τr≤T)≤yTχ(r)2.P(\tau_r\le T)\le\frac{y_T}{\chi(r)^2}.

For HdH^d, fixed v>0v\gt0 gives βdet>1\beta_{\rm det}\gt1, hence a>1a\gt1 for sufficiently small ε\varepsilon. Therefore

P(\tau_r\le T)\le\frac{T\varepsilon^2}{r^2}\quad(\mathbb R^d).$$ For $H^d$ with $a\gt1$, $b\gt0$ and a target probability $p_{\rm tar}\in(0,1)$, the bound $P(\tau_r\le T)\le p_{\rm tar}$ holds for every integer horizon $$T\le\frac{\ln[1+p_{\rm tar}(a-1)\chi(r)^2/b]}{\ln a}.$$ **Numerical checks.** Unless stated otherwise, $d=2$, $v=0.3$, and the noise has fixed norm $\varepsilon=0.01$ and uniform direction. - Quadrature gives $a=0.956286102922$, $b=9.8517066233\times10^{-5}$ on $S^2$ and $a=1.046418607299$, $b=1.0151896470\times10^{-4}$ on $H^2$. Direct integration over both action and noise directions agrees with the affine identity for tested separations up to $2.8$ on $S^2$ and $5$ on $H^2$. - On $H^2$, for $r=0.5$ and $T=20,40,60,80$, the exit bounds are 1.3%, 4.4%, 12.2%, 31.5%. The observed frequencies in 200,000 rollouts are 0, 0, 0.035%, 4.5%. For $p_{\rm tar}=0.05$, the formula guarantees $T\le42$; with $\varepsilon=0.001$ instead, it guarantees $T\le140$. - On $H^2$, at $t=40,100,150,200,300,400$, exact predicted RMS chordal error divided by simulated RMS chordal error is $1.0001,1.0006,1.0005,0.9999,1.0000,1.0003$ (200,000 rollouts). The paper's reduced-recurrence RMS prediction divided by simulated geodesic RMS error is $1.00,1.02,1.12,1.57,6.06,35.6$, respectively. - For the Table 3 protocol on $H^2$, the base step is $0.1$, $v_{\rm eff}=0.1H$, and $T=150$ decision events. As in Table 3, the predicted and simulated RMS chordal errors are each averaged over the final 30 events; their ratio is $1.000,1.001,0.999,1.001,1.008$ for $H=1,2,3,5,8$ (200,000 rollouts each). - On $S^2$, the exact chordal RMS floor is $\sqrt{b/(1-a)}=0.0475$. The simulated geodesic RMS error is $0.0478$ (1,000 rollouts of 2,000 steps).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.