- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes Geometry-Informed Planning (GIP) for offline goal-conditioned reinforcement learning. The central idea is to view the state space as a smooth manifold equipped with a temporal quasimetric, such that optimal goal-reaching trajectories correspond to temporal geodesics. Motivated by the exponential-map representation of geodesics, GIP represents an entire goal-reaching path using a single fixed-dimensional latent path vector. The method jointly learns a temporal distance function, a path map, a latent path-vector selector, and a local controller. The distance-learning objective additionally penalizes generated paths through unsupported regions, with the aim of discouraging invalid shortcuts.
The paper provides theoretical results characterizing zero geodesic cost, asymptotic path recovery under idealized assumptions, and an out-of-support penalization mechanism. Empirically, GIP is evaluated on OGBench across locomotion and manipulation tasks, where it achieves strong average performance, particularly on long-horizon humanoidmaze tasks. The paper also reports substantially lower inference cost than test-time search methods at comparable performance, together with scalability experiments, qualitative path visualizations, and several ablations.
Critical strengths
- Novel and practically meaningful formulation of long-horizon planning. The strongest contribution is the idea of amortizing an entire goal-reaching plan into a single fixed-dimensional latent path vector. Compared with explicit graph search or horizon-length trajectory generation, this provides a conceptually different planning interface whose inference cost need not scale directly with the planning horizon. This is both technically interesting and potentially important for real-time long-horizon control.
- Strong empirical results, particularly on challenging long-horizon tasks. GIP achieves strong performance on OGBench and is particularly competitive on the high-dimensional humanoidmaze tasks. The reported inference-efficiency gains are also substantial: on the humanoidmaze-giant settings, the paper reports 25.4× lower inference FLOPs than HIQL+TTGS at comparable performance. Together with the larger-scale colossal experiments, these results provide meaningful empirical support for the proposed amortized-planning formulation.
Critical weaknesses
- The connection between the strongest theoretical guarantees and the implemented algorithm could be made more precise. The path-recovery and distance-recovery results rely on idealized assumptions, including access to the true temporal distance, on-manifold paths, realizability, and global optimization, while the practical method does not directly guarantee these conditions. In addition, the learned encoder may not preserve the identity property required by some quasimetric arguments.
- The experimental study does not yet fully isolate the contribution of the core geometric components. While the paper includes ablations on latent dimension and state representation, it lacks targeted ablations for key ingredients such as the constant-speed geodesic objective, out-of-support penalization, joint distance learning, and the selector regularization. This makes it difficult to determine which parts of the proposed geometry-informed formulation are responsible for the observed gains.
Questions for authors
- The theoretical analysis uses the identity property , whereas the practical distance is obtained by applying an MRN quasimetric to learned representations and . Is there any mechanism that guarantees or approximately enforces injectivity of ? If not, how should the theoretical results be interpreted when distinct states may share the same learned representation?
- The out-of-support analysis assumes a strictly increasing path-length objective, while the implementation uses a clipped surrogate that stops applying additional spreading pressure after reaching a threshold proportional to . Does an analogous OOS exclusion argument hold for the actual surrogate? Also, since the practical method uses only intermediate points, do you observe paths crossing unsupported regions between sampled anchors, and can you quantify this effect?
- Could the authors provide matched ablations removing, individually, (i) the constant-speed geodesic loss, (ii) the generated-path/OOS term, (iii) joint distance learning, and (iv) the proximity/multimodal selector regularization? These results would materially strengthen my assessment by showing which components are responsible for the empirical gains.
- Can the authors provide a more controlled quantitative comparison with the strongest planning/trajectory-modeling baselines, such as CompDiffuser?
5 comments
Response to the review (2/3)
- Ablations of removing (i) the constant-speed geodesic loss, (ii) the generated-path/OOS term, (iii) joint distance learning, and (iv) the proximity/multimodal selector regularization.
We first provide ablation results obtained by individually removing the constant speed term, path length maximization term for OOS penalization, and proximity regularization term from GIP training.
Table 3. Ablation of geometric objectives (4 seeds).
| Environment | Full | w/o constant speed | w/o path-length maximization | w/o proximity regularization |
|---|---|---|---|---|
humanoidmaze-large-navigate | 76.6 ± 0.5 | 11.1 ± 5.2 | 67.7 ± 6.0 | 76.7 ± 2.2 |
humanoidmaze-giant-navigate | 81.5 ± 2.8 | 1.8 ± 1.5 | 36.6 ± 10.9 | 81.7 ± 1.3 |
The constant speed condition is essential to temporal geodesics, as established in the definition of geodesics in Eq. (3) and Proposition 4.2. Without this term, the path model fails to learn temporal geodesics properly, resulting in a severe performance drop. Without path length maximization, distances along paths passing through regions outside the state manifold can be underestimated, causing such paths to be incorrectly regarded as valid and leading to infeasible plans (e.g., paths crossing walls in maze environments).
Although proximity regularization has little effect on OGBench performance, we observed in the cylinder example that removing it causes premature mode collapse, limiting early exploration and preventing the overall geodesic cost landscape from being learned properly. We have added these results in Figure 8 of Appendix F.1. The ablation results are presented in Figure 10 of the revised manuscript, with further discussion in Appendix F.4 (line 1875).
To examine the effect of joint distance learning, we first jointly trained the distance model and temporal geodesic for 200k steps, which is sufficient for distance convergence, and then froze it while continuing temporal geodesic learning. The results are shown below.
Table 4. Ablation of joint distance learning (4 seeds).
| Environment | Joint distance | Frozen distance |
|---|---|---|
humanoidmaze-large-navigate | 77.1 ± 1.5 | 65.2 ± 8.7 |
humanoidmaze-giant-navigate | 81.1 ± 2.6 | 65.2 ± 6.0 |
Without joint distance learning, distance regression and OOS penalization using the intermediate points generated by temporal geodesics cannot be performed effectively, resulting in reduced performance. We have added these results as Figure 11 and included the corresponding discussion in Appendix F.4 (line 1886).
Response to the review (3/3)
- Comparison to CompDiffuser.
We note that the main AntMaze and HumanoidMaze experiments of CompDiffuser use xy-coordinate planning, whereas GIP generates full-state paths. Moreover, for HumanoidMaze medium, large, and giant, CompDiffuser uses maximum evaluation episode lengths of 5,000, 5,000, and 8,000 steps, respectively, compared with 2,000, 2,000, and 4,000 under the standard OGBench protocol followed by GIP. These differences prevent a fully controlled comparison of the reported scores. Furthermore, training and evaluating CompDiffuser in high-dimensional settings such as HumanoidMaze would require approximately one month with our available computational resources, which was not feasible during the rebuttal period.
Nevertheless, we compared GIP with the published results of CompDiffuser.
Table 5. Comparison of GIP (full-state planning) with CompDiffuser (xy-coordinate planning). CompDiffuser results are taken from Table 2 of the original paper (5 seeds); GIP results are taken from our full OGBench results (8 seeds).
| Environment | CompDiffuser | GIP |
|---|---|---|
antmaze-medium-stitch | 96 ± 2 | 86 ± 6 |
antmaze-large-stitch | 86 ± 2 | 78 ± 3 |
antmaze-giant-stitch | 65 ± 3 | 77 ± 6 |
antmaze-medium-explore | 81 ± 2 | 87 ± 4 |
antmaze-large-explore | 27 ± 1 | 71 ± 5 |
humanoidmaze-medium-stitch | 91 ± 1 | 91 ± 2 |
humanoidmaze-large-stitch | 72 ± 3 | 67 ± 2 |
humanoidmaze-giant-stitch | 67 ± 4 | 81 ± 5 |
| Average | 73.1 | 79.8 |
The results in Table 5 suggest that GIP achieves comparable goal-reaching success while generating full-state paths and operating under the shorter standard evaluation limits, which are more restrictive than those used by CompDiffuser.
CompDiffuser also reports results on antmaze-stitch tasks using full 29-dimensional state planning. Table 6 compares these published full-state results with GIP.
Table 6. Comparison of GIP (full-state planning) with CompDiffuser (full-state planning). CompDiffuser results are taken from Table 4 of the original paper (5 seeds); GIP results are taken from our full OGBench results (8 seeds).
| Environment | CompDiffuser (29D) | GIP (29D) |
|---|---|---|
antmaze-medium-stitch | 97 ± 2 | 86 ± 6 |
antmaze-large-stitch | 66 ± 5 | 78 ± 3 |
antmaze-giant-stitch | 28 ± 4 | 77 ± 6 |
| Average | 63.7 | 80.3 |
Although GIP achieves a lower success rate than CompDiffuser on the medium environment, CompDiffuser exhibits a severe performance drop as the task horizon increases. In contrast, GIP shows better scalability with respect to the horizon, as its latent path vector has a fixed dimension independent of the planning horizon. Consequently, the performance gap becomes larger as the environment size increases.
Another notable difference is the per-step inference FLOPs. Averaging the inference FLOPs per environment step across the available OGBench tasks in Table 1 of our manuscript, CompDiffuser requires approximately 5,000 times more computation than GIP. This demonstrates that GIP's implicit planning achieves remarkable computational efficiency while retaining the benefits of planning for long-horizon tasks.
We did not include CompDiffuser in the main result tables (Tables 1 and 5 of the manuscript) to ensure a fair comparison with other baselines, since CompDiffuser primarily plans in xy coordinates and its reported results were evaluated under more relaxed episode limits. However, we have added CompDiffuser to Figure 1 to allow an overall comparison of efficiency and performance.
Response to the review (1/3)
We thank the reviewer xqMf for the detailed feedback on the practical implementations and experiments, which helped us clarify our implementation details. Below, we address the reviewer's concerns and questions.
As you pointed out, the MRN architecture does not satisfy the identity property. Since we did not include the MRN distance formulation in the main text, we expressed the endpoint boundary term in Eq. (7) (line 238) as $ \rho(d^{*}(x_1,g)) $ for simplicity of presentation. In our implementation, however, we adopt a practical endpoint residual, $ \lVert \varphi_\phi(x_1) - \varphi_\phi(g) \rVert_2 + \lambda_\mathrm{state} \lVert x_1-g \rVert_2 $, which combines the norm between the endpoint and goal MRN representations with the norm between their raw states. We do not impose any injectivity on the encoder.
A strictly positive λstate ensures that a zero endpoint residual implies equality of the endpoint and goal states. We also conducted an ablation study on λstate.
Table 1. Ablation of λstate in the endpoint residual (4 seeds).
humanoidmaze-large-navigatehumanoidmaze-giant-navigateInterestingly, adding the raw state term did not yield a consistent improvement in performance. We therefore use only the representation norm in our final implementation. This is an empirical design choice: the final endpoint residual guarantees equality of representations, but not necessarily equality of the original states.
We had omitted these implementation details from the original manuscript, and we thank the reviewer for pointing out this omission. We now added a detailed description of the endpoint residual implementation, along with the ablation results, in Appendices D.4 and F.4 (lines 1332 and 1814).
An analogous argument holds under the assumptions of the original OOS analysis. It is important to note that τmin>1. Zero expected temporal geodesic cost requires sampled path length ℓϕ=∑k=0K−1dϕ(xk,xk+1) to be exactly dϕ(s,g), whereas clipping occurs at the larger value τmin∗dϕ(s,g).
With generated points and targets fixed, the distance modification in the original proof increases the length of a path containing sampled OOS points while preserving dϕ(s,g). Starting from a zero-cost path, this both reduces the clipped distance loss and makes the path length greater than dϕ(s,g), ruling out zero geodesic cost under the modified distance. The expected loss decreases strictly when such OOS points are sampled with positive probability. Clipping stops further spreading only after the path already violates the zero-cost condition, so the clipped loss still eliminates possibility of zero-cost OOS paths.
First, we would like to clarify a detail that we missed in the original manuscript. In Eq. (11), the interior times t1,t2,…,tK−1 are sampled from Unif[0,1] at each training iteration. We noticed that this was not explicitly stated and have clarified it in the revised manuscript (line 283):
... 0=t0<⋯<tK=1 are the discretization times, with interior times resampled from Unif[0,1] at each iteration, ...
If the time points were sampled only once with K=4, a path could cross an unsupported region between the sampled points. We observe such crossings, particularly early in training. However, since the interior times are independently resampled at every training iteration, repeated sampling allows the entire interval [0,1] to be probabilistically covered over the course of training.
Since the time points are resampled at each iteration, it is difficult to directly quantify violations as a function of K. Instead, we conducted an ablation study on K.
Table 2. Ablation of K (4 seeds).
humanoidmaze-large-navigatehumanoidmaze-giant-navigateThe results show that increasing K to 8 yields a modest improvement in performance, but naturally increases training time as well. Considering this tradeoff between efficiency and performance, we chose K=4 as a reasonable balance.