acceptodds
Under review as a conference paper at ICLR 2027

Relative Flow Matching for Time-Critical Control in Hardware-Limited Flight

Abstract

Generative policies are a natural fit for multimodal robot control, but generative sampling can be too expensive for time-critical control on hardware-limited platforms. We introduce Relative Flow Matching (RFM), a reparameterization of flow matching that moves the fixed distribution source towards an action-dependent anchor. This is particularly useful in robotics, where information on prior actions can substantially reduce the transport required between source and consecutive action. By defining the flow in relative coordinates, RFM retains a standard Gaussian source without loss of generality, allowing existing flow-matching objectives and fast samplers to be applied directly. We evaluate RFM on collision-free flight of a morphing-wing drone, where RFM reaches % success in simulation while requiring ms per inference on a Raspberry Pi Zero 2W, and transfers to the physical aircraft with 14/15 successful flights. We show that RFM outperforms conventional Flow Matching and other action conditioned methods in terms of performance and speed. Across standard Robomimic and D3IL simulated manipulation benchmarks, RFM improves one-step generation in four of five settings and remains competitive with iterative flow matching at substantially lower inference cost. This work, establishes a continuous connection between standard fixed-source flow matching, stochastic action anchoring, and the deterministic behaviour-cloning boundary. Our results show that RFM's affine reparameterization materially improves very-low-NFE generative control because it changes the conditional source/transport geometry.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Flow Matching, Robotics, UAV, Drone

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_t7VYOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents Relative Flow Matching (RFM), which centres the flow-matching source on the previous action so that the network predicts a scaled residual rather than the action itself. Since this is essentially a change of variables, the theoretical contribution is limited. The application is the main strength. A one-step policy matches the MPC expert at 50 Hz on a Raspberry Pi and is validated on a real drone, with a clean comparison against MeanFlow. However, the paper's own analyses only partly support its explanation.

Critical strengths

The application result is strong and practical. A one-step policy reaches MPC-level success (95.2%) in 4.7 ms on a Raspberry Pi Zero, well within the 50 Hz control budget.

Critical weaknesses
  1. The method's novelty is limited. RFM amounts to an invertible affine change of variables, equivalent to standard flow matching on the scaled residual from the previous action, and its propositions follow directly from this. The results for A2A and Streaming Flow Policy are not reported in the main text, so it is hard to judge how much RFM improves on these closest methods.

  2. The paper's explanation of why RFM works is not supported by its own analyses. Calibrating σ\sigma without any anchor already gives much of the transport reduction and changes success on its own, yet the anchor and the σ scaling are never ablated separately.

Questions for authors
  1. What are the success rates and latencies of A2A and Streaming Flow Policy on the drone task, and how do they compare with RFM-MF under matched architecture and number of network evaluations?

  2. During training, is x_prev the expert's previous action chunk or a chunk generated by the policy? If it is the expert's, how large is the train-test mismatch at deployment?

4 comments
AuthorsOpenReview ↗

Response to Reviewer t7VY

Response to Reviewer t7VY

Please see also our general response, which covers the points shared across reviews.

We thank Reviewer t7VY for the clear review and for recognising the application result as strong and practical.

W1/Q1 Novelty, A2A and SFP in the main text; matched NFE. Our method defines a continuous interpolation s between standard FM and direct action mapping (A2A) in a residual space, see the general response for more detail. This differs from the previous works discussed in the paper that does not operate with residual flows (thus losing generality with fixed source methods). We find this residual space does not only provide a theoretically sound connection, but also shows 1) Improvement over direct action mapping in absolute space; 2) improved performance at an intermediate anchoring strength s (i.e., in between FM and A2A).

Both A2A and SFP are shown in main text figure 2. We will mention these values in the corresponding results section:

  • A2A: 0.94M, NFE = 6, 18.4 ms, 81.1 ± 2.1 %
  • SFP: 0.56M, NFE = 4, 11.2 ms, 76.8 ± 4.1 %
  • RFM-MF small: 0.96M, NFE = 1, 4.7 ms, 95.2 ± 1.8 % <- matched architecture size
  • RFM-MF big: 2.89M, NFE = 1, 10.0 ms, 95.9 ± 1.6 % <- matched inference time

Each baseline uses the architecture of its reference implementation (App. B), with A2A's latent width selected on our task (bigger latent had worse performance). For matched NFE, we evaluated the published A2A weights with a single Euler step: 83.8 ± 4.2 %. At the same single call and practically the same size (0.94M vs 0.96M), RFM-MF leads A2A by 11 points, so the gap is not a matter of step count; six steps remains A2A's intended procedure and its entry in the table. SFP takes one network call per action, so for the single executed action it already runs at one call; at matched NFE RFM-MF leads it by 18 points.

W2 Anchor and σ never ablated separately. Sigma and s were ablated (Figure 4 with full ablation reference to table 5); now also updated and shown in the general response. Rescaling without the anchor gives 77.0 % against MeanFlow's 78.7 %; the anchor in absolute action coordinates 84.8 %; the anchor in RFM's residual coordinates 92.5 %, and 95.9 % in the chained form used for all results. The σ columns of Table 5 give the response to the scale at fixed weights, and the measured σ* is the best scale in every anchored row. This shows that several factors result in the performance gain, with a combined effect that outperforms overall.

Q2 Expert's or policy's x_prev; size of the mismatch. The anchor is the expert's previous action during training and the policy's own previous command at deployment (App. B); no anchor noise is added. Every success rate in the paper is closed loop with the policy's own anchors, so the effect of this mismatch is already contained in the reported numbers. To measure its size, we ran the NMPC as a shadow expert on the states visited by RFM-MF over the 400 held-out episodes. The policy's anchor lies 0.20 (RMS, normalised action units) from the expert's command at the same state; with s = 0.5 half of this enters the source, well within the source's own standard deviation (σ* = 0.276). The residuals the policy generates in closed loop have RMS 0.235 against 0.276 in the demonstrations, so they stay inside the training scale rather than drifting outward. This is consistent with the role of s: with s < 1 an anchor error is contracted rather than passed on in full, and s = 0.5 outperforms s = 1 (Table 5).

Thank you for the author's reply. I maintain my score due to the limited contributions.

We thank the reviewer for the reply and for the effort they spent reviewing our paper.

Thank you for your reply, I appropriate the contributions in hardware experiments. But I strongly recommend authors submit to ICRA or IROS.

Reviewer_JKXnOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies low-latency generative control for a morphing-wing aircraft. Relative Flow Matching (RFM) replaces a fixed action-space source with x_0 = s x_prev + sigma epsilon and generates normalized residuals delta = (x - s x_prev) / sigma. For positive sigma, this is an invertible affine change of coordinates with a standard Gaussian residual source. The authors apply existing flow-matching objectives and MeanFlow in these coordinates, with the latter providing one-network-evaluation control. The theoretical results establish affine equivalence and characterize a deterministic-source boundary under the stated assumptions; they do not establish a general finite-step accuracy improvement.

The main flight simulation uses 10 independently trained policies per method and 400 held-out initial conditions per policy. A 0.96M-parameter RFM-MF model reports 95.2 +/- 1.8% success and 4.7 ms per policy call measured on a Raspberry Pi Zero 2 W. The larger model and fixed-source MeanFlow share a trunk and report 95.9 +/- 1.6% versus 78.7 +/- 4.9% success at the same 10.0 ms latency. Physical validation reports 14 successful flights out of 15 using motion-capture feedback, offboard computation, lateral PID stabilization, and a virtual obstacle. Five state/image manipulation settings, source-parameter sweeps, transport measurements, and a two-mode toy study provide additional evidence. The manipulation comparisons use some task-specific configurations and are explicitly not fully matched source comparisons.

Critical strengths

S1. A practically meaningful performance/latency result with a useful matched comparison. Meeting a 20 ms control period on a small embedded computer is a concrete constraint. The most persuasive comparison is the 2.89M-parameter RFM-MF versus MeanFlow on the drone: the manuscript reports the same architecture, inputs, NFE, and latency, with success increasing from 78.7 +/- 4.9% to 95.9 +/- 1.6% (Table 3). Unlike a comparison against only many-step diffusion, this supports a benefit beyond reducing NFE. Ten training seeds and separate checkpoint-selection and reporting conditions strengthen this result. The physical flights also show that the learned controller can transfer to an aircraft, within the offboard setup actually tested.

S2. A simple reusable formulation and analysis that exposes important limitations. The positive-sigma coordinate transformation is straightforward and compatible with established flow-matching machinery. Separating anchor strength from residual stochasticity is useful for studying fast control policies. The paper reports that shorter transport does not monotonically improve success, that severe disturbances can remove the advantage, and that a stochastic source alone does not guarantee multimodal outputs from a one-step sampler. These observations are more informative than reporting only successful trajectories. Manipulation results suggest potential beyond the aircraft, although they do not isolate the source change as cleanly as the drone comparison.

Critical weaknesses

W1. The distinction from prior action-conditioned flows and the attribution of the gain to the proposed reparameterization are not sufficiently established. Proposition 1 establishes equivalence to standard conditional flow matching after an affine transformation. Consequently, the contribution must rest mainly on the practical effect of the source/normalization choice rather than increased expressive power or a new flow-matching objective. The paper acknowledges this in its discussion, but its framing of prior work is too broad. In particular, Streaming Flow Policy already uses a narrow Gaussian around the previous action and explicitly models multimodal per-timestep distributions; this is not equivalent to a source that has collapsed to a deterministic point. Its limitations concern joint versus marginal trajectory distributions and its particular execution scheme. These distinctions are stated in the primary paper (Streaming Flow Policy, Sections 2, 4, and 9). Proposition 2 concerns an exactly deterministic source and, for the conditional-mean statement, a particular one-step squared-error prediction. It does not establish that existing finite-variance anchoring methods lose multimodal expressivity or that RFM-MF preserves it at one NFE.

The matched MeanFlow comparison changes both source location and scale. Table 7 shows that much of the measured transport reduction already occurs at s = 0 with a calibrated scale: on the drone, the distance drops from 4.92 for N(0,I) to 1.75 at the unanchored calibrated source, and then to 1.16 at s = 0.5. The sweeps are useful, but a clearly labeled matched comparison among standard MeanFlow, scale-calibrated unanchored MeanFlow, anchored action-space MeanFlow, and the normalized residual implementation would make the contribution much easier to assess. The last comparison may be equivalent under an appropriate network transformation; that equivalence itself would clarify what is a source-design contribution and what is an implementation/optimization benefit. Current evidence supports the usefulness of the combined recipe more strongly than the claimed distinction from prior action anchoring.

W2. Important implementation and reporting inconsistencies limit reproducibility and the scope of the deployment conclusion. Section 3 defines a previous-chunk affine transformation with an anchor fixed by the conditioning. Appendix B instead describes the latest executed action as the initial anchor and chained within-chunk increments reconstructed cumulatively in Eq. 12. A chunk-wise translation and a cumulative-increment representation are different transformations. Proposition 1 does not directly specify the latter, and changing the chunk representation could affect the purported source-only control. The manuscript needs one consistent algorithm, including the exact role of the 20 history inputs, the training targets, sampling reconstruction, and the representation used by the matched MeanFlow baseline. The execution protocol also needs clarification: Section 4.1 describes H = 4 and executing only the first action, while Appendix B specifies DDPM chunk 16 / execute 8.

Several numerical details also obstruct verification of the proposed mechanism. The discussion quotes 68.5 +/- 3.4% for the matched drone MeanFlow comparison, whereas Section 4.2 and Table 3 give 78.7 +/- 4.9%. Table 5 lacks the sigma-multiplier column labels in the supplied PDF, calls its entries single-run sensitivity while the surrounding text refers to 10 seeds, and does not make it possible to identify and reconcile the nominal s = 0.5, sigma = 0.276 configuration with the 95.9 +/- 1.6% main result. These issues directly affect interpretation of the key source ablation.

Finally, Appendix B states that both aircraft controllers ran on an external laptop and that lateral attitude was stabilized by PID. The physical obstacle was virtual, and the 4.7 ms figure is a separate Raspberry Pi timing measurement. This is useful evidence of transfer plus embedded inference feasibility, but it does not demonstrate an integrated onboard perception-to-actuation loop meeting its deadline. The front matter should distinguish those claims consistently. The offboard setup is disclosed, so this is a claim-scope concern rather than an allegation of misconduct.

Questions for authors
  1. Which transformation produced the reported results: the fixed previous-chunk affine map in Section 3, or the chained increments in Appendix B, Eq. 12? Please provide concise training/inference pseudocode and state whether the matched MeanFlow model uses the same chunk representation and action-history conditioning. Also reconcile the stated execute-one protocol with DDPM chunk 16 / execute 8.
  2. Please restore all Table 5 column labels, define its uncertainty and seed count, identify the nominal configuration, and reconcile its values with Table 3. Which MeanFlow figure is correct for the main matched comparison: 78.7 +/- 4.9% or 68.5 +/- 3.4%?
  3. What exactly is claimed to be unavailable to existing non-degenerate Gaussian anchoring methods? Please distinguish full-rank source expressivity, one-step approximation quality, and Streaming Flow Policy's marginal/joint-distribution distinction. Is the controlled two-mode one-step result obtained by a single Euler step of an instantaneous flow field, rather than a separately trained MeanFlow model? If so, please limit the corresponding conclusion to that tested sampler.
  4. Please confirm the real-flight control frequencies, end-to-end sensing/communication/actuation latency, and which components, if any, ran onboard. State the Raspberry Pi inference benchmark and the offboard flight demonstration as separate results. Fifteen flights do not establish hardware reliability or an advantage over the expert's 15/15 result.
  5. Using the existing drone setup and matched MeanFlow architecture, report a clearly specified comparison of standard fixed-source generation, unanchored scale-calibrated generation, and anchored residual generation, with seed-level uncertainty and the same checkpoint-selection protocol. Include an anchored action-space implementation if it is meaningfully different from the residual implementation; otherwise explain their equivalence. This would address the main contribution-attribution issue without requiring a new benchmark.
5 comments
AuthorsOpenReview ↗

Response to Reviewer JKXn, part 1

Response to Reviewer JKXn

Please see also our general response, which covers the points shared across reviews.

We thank Reviewer JKXn for the thorough review, and detailed response.

W1a RFM distinction with SFP/A2A; scope of Proposition 2. SFP trains with a narrow Gaussian around the previous action, but at test time sets σ₀ = 0 "in all our experiments" (SFP §4, deterministic execution); its latent variant uses σ₀ ≈ 0 (SFP App. B). A2A starts from a deterministic encoding of the action history (A2A §3.2). Both are unable to achieve good performance with a narrow gaussian in absolute action space (as described in A2A §5.2), thus therefore deployed from a deterministic anchor, the s → 1, σ → 0 end of our parameterisation, which is what the introduction and §3.1 describe. We found the same performance drop when applying A2A/SFP and MF in absolute space vs relative. This indicate that fixed source FM and anchoring methods cannot be straightforwardly combined, which our method solves through its residual Flows.

We make will explain this more explicit: SFP trains with a stochastic source, but executes deterministically (like A2A), and matches per-timestep marginals rather than the joint distribution of the chunk. Proposition 2 concerns the σ = 0 boundary only. RFM-MF instead keeps its finite source at deployment, and on our task this is worth 10 points: the same weights sampled near-deterministically (0.1σ*) reach 85.2 %, at the calibrated σ* 95.8 % (Table 5, s = 0.5).

W1b/Q5 Four-way comparison. We agree this is an interesting analysis and will add it in the revision, the full table is in the general response (ten seeds each). The results support your idea on a combined recipe for the performance found. Rescaling alone shortens the transport (4.92 → 1.75) without raising success at fixed N(0, I) 78.7 %, rescaled without anchor 77.0 %. The gain requires the anchor (obtains 84.8 %) in absolute action coordinates and is completed by expressing it in residual coordinates 92.5 %, and 95.9 % as RFM-MF chained actions. So transport length alone is predictive of performance as the ablation shows.

W2a/Q1 One algorithm; history inputs; MeanFlow; horizons. All RFM-MF results use Eq. 12: the anchor is the single action preceding the chunk, and each action in the chunk is generated relative to its predecessor. Eq. 12 is an invertible affine map of the chunk, x = s·x_prev + σLδ with L the cumulative-sum matrix, so Proposition 1 covers it directly; the revised §3.1 states the method in this form, with an algorithm box. The 20 action dimensions of the 41-D input hold the four most recently executed actions not the previous chunk; only the most recent action acts as the anchor. The matched MeanFlow uses the same trunk and input and predicts the absolute chunk from N(0, I). DDPM and Consistency Policy use chunk 16, following their reference recipes; trunk, chunk length, executed actions and NFE of every row are listed in table 3 (App B).

AuthorsOpenReview ↗

Response to Reviewer JKXn, part 2

PSEUDO CODE

 x_prev ← executed action preceding the chunk                 # most recent entry                                                                                                                                                                                                                    
 μ      ← s · x_prev                                          # s = 0.5                                                                                                                                                                                                                                                 
 δ₁     ← [(a₀ − μ), (a₁ − a₀), …, (a_{H−1} − a_{H−2})] / σ   # Eq. 12, σ = 0.276                                                                                                                                                                                                                                 
 ε      ~ N(0, I)                                                                                                                                                                                                                                                                                                 
 (r, t) ~ MeanFlow time sampler                                                                                                                                                                                                                                                                                   
 z      ← (1 − t) ε + t δ₁                                                                                                                                                                                                                                                                                        
 loss   ← MeanFlow objective on u_θ(z, r, t | c) with velocity δ₁ − ε   # Geng et al.,                                                                                                                                                                                                            
                                                                                                                                                                                                                                                                                                               

Inference:

 μ   ← s · x_prev                                                                                                                                                                                                                                                                                                 
 δ   ← ε + u_θ(ε, 0, 1 | c),   ε ~ N(0, I)                                                                                                                                                                                                                                                                        
 a_k ← μ + σ Σ_{j≤k} δ_j        for k = 0…H−1   # x = s·x_prev + σLδ                                                                                                                                                                                                                                              
 execute a₀ (clipped to [−1, 1]); x_prev ← a₀; append a₀ to the action history       

W2b confusing numbers We apologies for our errors that occured in the last minutes before the deadline. We will correct them in the revision:

  • 68.5 vs 78.7, 78.7 ± 4.9 % is the correct value and used throughout, 68.5 % was a stale value from a shorter schedule and is corrected.
  • Table 5 The submitted grid accidentally contained the small model. We re-ran the full grid on the 2.89M, every cell is the mean (N=10) with 95 % CI on 400 held-out conditions.
sσ*0.1σ*0.25σ*0.5σ*1σ*2σ*4σ*
0.00.49784.8±0.384.3±0.785.7±3.077.0±4.562.6±4.749.5±15.4
0.30.36085.2±0.484.5±1.187.1±2.391.1±1.470.4±2.673.9±4.2
0.50.27685.2±0.485.0±1.189.8±1.295.8±1.373.2±2.475.2±3.2
0.70.20684.2±0.684.7±1.492.9±1.595.2±1.761.1±5.952.7±5.2
1.00.16679.3±3.777.5±8.782.9±10.883.9±9.853.7±7.459.1±5.5
AuthorsOpenReview ↗

Response to Reviewer JKXn, part 3

W2c/Q4 Real Flights. See the deployment paragraph of the general response. RFM-MF ran at 50 Hz it was trained on, while the NMPC expert could only reliably reach 30 Hz on our machine (H=30-steps). Both ran on the same laptop to make the overhead consistent (MoCap -> Laptop: inference -> send action commands -> drone). Both used the same lateral PID for heading and roll, with a virtual beam for the NMPC's constraints (the policy does not sense the beam). The fifteen flights are reported as evidence of transfer, not as a reliability estimate, which the paper does not claim.

Q3 What finite-variance anchoring lacks; toy sampler We never claim gain through expressivity, which RFM (see general response).

  • Full-rank source expressivity, Flow matching from any non-degenerate Gaussian source can reach the target, and RFM is an invertible affine reparameterisation of such a source (Prop. 1). We make no expressivity claim against finite-variance methods; the paper's multimodality statements concern only σ → 0 (Prop. 2).
  • One-step approximation quality, this is RFM's performance claim. The two parameterisations share the same optimum (Prop. 1) but pose different regression problems to a one-step network: residual coordinates remove the conditioning-dependent offset s·x_prev and put source and target at unit scale for every anchor (Cor. 1b). With the same anchored source, trunk and single call, MeanFlow reaches 84.8 % in absolute action coordinates and 92.5 % in residual coordinates (95.9 % with the chained residuals of Eq. 12; general response). Anchoring alone recovers part of the gain over the fixed source (78.7 → 84.8 %); the residual coordinates recover the rest. A2A's own ablation finds its flow converges worse in raw action space than in its latent space (A2A §5.2).
  • SFP's marginal/joint distinction SFP matches per-timestep marginals, while RFM's flow is defined on the whole chunk and targets its joint distribution. This is a modelling difference; the ablation attributes the drone result to the anchor and the residual coordinates. Furthermore, at test-time SFP's source is degenerate, resulting in a point mass, so for a given history it returns a single trajectory, the setting of the first part of Proposition 2.

Toy sampler. Yes, the one-step row was a single Euler step of the instantaneous field. We re-ran it with a separately trained MeanFlow model in residual coordinates, the sampler RFM-MF deploys (three seeds per σ). It stays on a mode for every σ ≥ 0.1 (d = 0.49–0.51, where 0.5 is a mode and 0 the mean; both modes drawn equally often) and approaches the mean only as σ → 0, the boundary of Proposition 2, while RFM-MF keeps both modes at one NFE for finite σ.

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

The paper presents Relative Flow Matching, which constructs a noisy source around the previous action and uses relative coordinates to formulate action generation as standard flow matching. The method is combined with MeanFlow and evaluated on drone obstacle avoidance and manipulation tasks, with an emphasis on control using very few network evaluations.

Critical strengths

The method is simple and easy to combine with existing generative policies. The drone experiments address a concrete computational constraint and include inference measurements on the target hardware. The comparison with MeanFlow using the same network architecture also provides useful evidence of the practical benefit.

Critical weaknesses

My main concern is the limited novelty. Starting generation from a noisy previous action is a natural choice for sequential control, and previous-action anchoring is already used by previous papers, as discussed in the paper. RFM's main addition is to subtract the anchor and divide by the noise scale, standardizing the conditional Gaussian source through an affine transformation. This does not increase model expressiveness. The equivalence result explains why existing training and sampling methods can be reused, but I am not convinced that this alone constitutes a sufficiently strong methodological contribution.

The experiments do not clearly establish which part of the proposed change produces the gains. The paper emphasizes shorter source-to-target transport, but adjusting the noise scale alone substantially reduces this distance, and shorter transport does not consistently improve success. There is also a difference between the fixed-anchor residual formulation in the main text and the within-chunk action differences with cumulative reconstruction described in the appendix. This makes it difficult to separate the effects of anchoring, scaling, and action representation. These factors should be isolated using the same representation, network, and training configuration.

The method also introduces anchor strength ss and noise scale σ\sigma, adding a task-specific tuning burden. Although the paper studies parameter sensitivity, it is unclear whether the same parameters or selection rule transfer across tasks, platforms, and control frequencies. If each setting requires a separate search, that validation cost should be included when assessing the method's practical value.

The deployment evidence is narrower than the main presentation suggests. The paper emphasizes real-time control on limited onboard hardware, but Appendix B states that the physical flights use an external laptop; the Raspberry Pi results are separate inference measurements. The manipulation comparisons also do not consistently match architectures and prediction horizons. These experiments demonstrate potential, but provide weaker support for onboard deployment and generality than the paper suggests.

Questions for authors
  1. Beyond affine standardization and integration with MeanFlow, what is the substantive new mechanism relative to previous-action anchoring? With identical action representations and training configurations, how does RFM compare with using the same conditional Gaussian source directly in action coordinates?

  2. Can the authors isolate anchor location, noise scale, and within-chunk action differencing to establish which changes account for the improvement?

  3. Can one pair of ss and σ\sigma, or one automatic selection rule, transfer across tasks? Please compare fixed parameters with task-specific tuning and report both the performance benefit and the validation cost of tuning.

3 comments
AuthorsOpenReview ↗

Response to Reviewer frUM

Response to Reviewer frUM

Please see also our general response, which covers the points shared across reviews.

We thank Reviewer frUM for the in-depth review. Both attribution questions are now answered by controlled runs.

W1/Q1 Novelty; the same source directly in action coordinates.

  • Novelty Our method defines a continuous interpolation s between standard FM and direct action mapping (A2A) in a residual space, see the general response for more detail. This differs from the previous works discussed in the paper that does not operate with residual flows (thus losing generality with fixed source methods). We find this residual space does not only provide a theoretically sound connection, but also shows 1) Improvement over direct action mapping in absolute space; 2) improved performance at an intermediate anchoring strength s (i.e., in between FM and A2A).

  • Identical action representation We ran the additional experiments comparison: RFM-MF's anchored source (s = 0.5, σ* = 0.276) with the network trained and sampled in absolute action coordinates reaches 84.8 ± 2.5 %, against 95.9 ± 1.6 % for RFM-MF (same 2.89M trunk, schedule and protocol, ten seeds each). The two parameterisations describe the same family of flows (Prop. 1). Without an anchor (s=0) they differ only by a constant rescaling, which the first and last layers absorb. With an anchor, the residual coordinates also subtract s·x_prev, an offset that depends on the conditioning; in absolute coordinates the network has to learn this offset itself, and with a single call that costs 8 points at the same single anchor (84.8 vs 92.5 %) and 11 against RFM-MF. This is in line with A2A's own ablation, that showed that its flow converges worse in raw action space than in its latent space (A2A §5.2).

W2/Q2 Anchor, scale, or within-chunk differencing. Eq. 12 is itself an invertible affine map of the chunk, x = s·x_prev + σLδ with L the cumulative-sum matrix, so Prop. 1 covers it unchanged. The revised §3 states the method in this form, which is the form behind all RFM-MF results.

The table in the general response investigate the effects of RFM-MF design. Without the anchor, neither rescaling (77.0 %) nor within-chunk increments (75.6 %) improves on MeanFlow (78.7 %). With the anchor, absolute coordinates reach 84.8 %, standardised residual coordinates 92.5 %, and the chained residuals of Eq. 12 95.9 %; the increments therefore add about 3 points, and only on top of the anchor.

W3/Q3 Tuning burden and transfer. There is no per-task search. s = 0.5 was chosen on the drone and used unchanged (we found its optimum is broad 95.8 % at s = 0.5 and 95.2 % at s = 0.7, Table 5).

σ* is not a tuned hyper-parameter but a expert data statistic we obtain before learning (the residual RMS of the demonstrations scaled at s). This is computed in one pass over the training data, so it transfers as a rule rather than as a value. In the revised Table 5 (grid in our reply to JKXn), this measured σ* is the best sampling scale in every anchored row.

W4 Deployment evidence. We present the deployment details in the general response. A full mocap->PI-actuator command runs at 10ms (well within 50 Hz), with 65ms delay. On manipulation most setups are matched, except for 1) ToolHang from state uses a smaller trunk (18.8M instead of 34.5M) with observation horizon 5 instead of 2, and 2) Square from images a shorter chunk (8, executing 2, instead of 16 and 8). In the other three settings RFM-MF uses the same configuration as MeanFlow and FM. The revised App. D names both cells. Manipulator inference speed is reported as well and show consistent inference reductions accross different deployments.

The additional experiments clarify the empirical benefits and address several of my concerns. However, I still view RFM primarily as a modular combination of previous-action initialization and the existing MeanFlow framework through an affine residual reparameterization, and I find the methodological novelty insufficient to support acceptance. I therefore keep my overall rating.

We thank the reviewer for the follow-up and are glad that the additional experiments address the empirical concerns. We would still like to emphasise that the central contribution of RFM is the formulation of anchoring in standardised residual coordinates, which makes it compatible with standard flow matching objectives and results in better performance than a straightforward modular combination of previous-action initialisation and MeanFlow.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

General response to all reviewers

We thank the reviewers for their reviews and are happy to clarify our contribution. We are encouraged for the recognition of RMF's practicality and elegant design, and the strength of the drone result, which fits the applications to robotics track. Three general questions we answer here: what differentiates RFM from an affine change of variables? Which part of RFM produces the gain in performance? Can we expand more on the real world deployment? We answer them here and go more in depth on the remaining points in the individual replies.

Contribution. As mentioned in the paper, we do not claim that RFM adds expressivity (§5). RFM contributes a previous-action anchoring method compatible with standard (one-step) flow training, that is theoretically sound by introducing a residual space. This residual space lets us train a one-step generative policy reach expert-level control inside a 20 ms control period for direct drone control.

  1. A continuation interpolation. x₀ = s·x_prev + σε spans fixed-source flow matching (s = 0, σ = 1) and previous-action anchoring (s → 1, σ → 0) with two interpretable parameters.
  2. Standard recipes apply unchanged. In residual coordinates the source is N(0, I) for every anchor (Prop. 1), so MeanFlow, and any recipe derived for a standard N(0, I) source, applies with its standard loss and a single network call. Prior anchoring methods do not address this source mismatch and execute from a deterministic anchor: A2A anchors a deterministic latent encoding, adds an autoencoder and an unrolled consistency loss, and its own ablation finds that the flow converges worse in raw action space than in that latent space (A2A §5.2); SFP trains with a narrow Gaussian but sets σ₀ = 0 at test time and matches per-timestep marginals rather than the joint distribution of the chunk. RFM keeps its calibrated stochastic source at deployment.
  3. Anchor and scaling contribute to the performance. Rescaling alone gives nothing without anchoring (RMF s=0 σ=σ*, 77.0 % against MeanFlow’s 78.7 %), while anchoring alone in absolute action coordinates, reaches 84.8 %; together, in RFM’s standardised residual coordinates, they reach 95.9 %. Reparameterisations that leave the model class unchanged but change what a network learns at fixed capacity, such as the preconditioning of EDM (Karras et al., 2022), matter most in exactly this regime of a single network evaluation.

Where the gain comes from. RFM-MF and the matched MeanFlow share the 2.89M trunk, data, 480-epoch schedule and one network call. The RFM construction has three parts: the anchor (s), the source scale (σ* instead of 1), and the coordinates the network works in (absolute actions, standardised residuals, or the chained residuals of Eq. 12 used for all RFM-MF results). We ran ran a combinatorial sweep and present the results below (drone task, ten seeds each, mean ± 95 % CI, protocol of Table 3):

sourceaction coordinatessuccess [%]
fixed N(0, I) (MeanFlow)absolute actions78.7 ± 4.9
rescaled, no anchor (s = 0, σ* = 0.497)standardised77.0 ± 4.5
fixed N(0, I), no anchorwithin-chunk increments75.6 ± 4.0
anchored (s = 0.5, σ* = 0.276)absolute actions84.8 ± 2.5
anchoredstandardised residuals, single anchor92.5 ± 1.2
anchoredstandardised residuals, chained (Eq. 12) = RFM-MF95.9 ± 1.6

Without the anchor, neither rescaling nor within-chunk increments improves on MeanFlow. The anchor in absolute action coordinates adds 6 points; expressing the same anchored source in residual coordinates adds a further 11. The two are complementary. Without an anchor the two coordinate systems differ only by a constant scale, which the first and last layers absorb; around an anchor the residual coordinates also remove the offset s·x_prev, which depends on the conditioning. Together they are the construction the paper proposes.

Deployment.

  • Unclear inference speed benchmark Current work presents the inference speed benchmark at the simulated drone results, separate from the real world validation. In the revision we will emphasize the deployment difference more explicit. Additional full closed-loop tests (mocap->Pi->drone, not in flight) to measure delay/inference speeds show about 10 ms control cycle, well inside the 20 ms period.
  • Closed loop in simulation: All success rates are 50 Hz closed-loop rollouts including communication/sensor/actuation delay (65 ms). The simulated environment, and expert(data) model this deployment overhead (witheld for anonymity).
  • Physical flights: 14/15 successful flights with the policy computed off board, on the same laptop as the NMPC expert, because the expert is unable to fly with the Pi we choose to standardize the overhead. This validates tranferability of the learned policy.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.