{"id":"418ab5f4-f7cb-443f-bcd9-317b6c3bf43f","arxiv_id":"2411.17293","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"SIL-RRT* trains a transformer-based sampler with self-imitation learning to guide RRT* tree expansion, reporting large sample-count reductions in 2D, 3D, and snake planning benchmarks.","lead":"SIL-RRT* is a path-planning algorithm that adds a learned 'where to sample next' model to the classic RRT* tree-search method, using a transformer network trained on its own past successes. It reports needing far fewer random samples than standard RRT* to find collision-free paths in 2D, 3D, and a simple 5-joint snake robot.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed speedups are confounded: SIL-RRT* embeds BiRRT* with asymmetric sample caps, and no ablation isolates the learned sampler's contribution.","rationale":"The reader's no-ablation point matches my main concern, but I would not single out the multivariate Gaussian family as the weakest link: the paper provides no formal capacity analysis, and the observed confound is more directly damaging because it undermines attribution of the results to any learned sampler. The most load-bearing issue is causal identification: SIL-RRT* changes both the sampling distribution and the search strategy, and the evaluation does not separate them. The current CONDITIONAL verdict is appropriate; this concern does not move it to rejection, but it strengthens the condition that a clean ablation must be provided. Secondary issues, such as only three trials, the duplicated Section 3.3/4 text, and the mismatch between SIL-RRT* surface point clouds and MPNetNR interior point clouds, further reduce confidence but are not needed to justify the conditional verdict.","tokens_in":9997,"tokens_out":3441,"duration_ms":34951,"concrete_test":"Run the 2D, rigid-body, and 3D test suites four ways: (1) SIL-RRT* as published; (2) SIL-RRT* with the learned sampler replaced by a uniform sampler, keeping BiRRT*, the same node cap, stopping rule, and goal-region definition; (3) unidirectional RRT* with the learned sampler; (4) plain RRT* as baseline. If (2) matches (1) within a small margin (e.g., success rate within 2 percentage points and samples within 10%), the headline speedup is due to the bidirectional search, not to the learned distribution. If (1) is substantially better, the central claim survives this confound. The same test should be repeated on the 3D environment with equal sample caps (e.g., 200 vs 200) to remove the cap asymmetry.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the learned Gaussian sampler, trained by self-imitation, reduces samples and time relative to RRT* (Section 5.2). The load-bearing condition is that the observed gains are caused by that learned distribution and not by other components introduced in the same algorithm. That condition is not established. Section 4.1 states that SIL-RRT* integrates BiRRT*, which alternates forward and backward tree expansions, explicitly to fix failures to find feasible paths; plain RRT* in Table 1 has no such bidirectional search. The 3D row also caps RRT* at 400 samples while SIL-RRT* is capped at 200, so 'fewer samples' is partly by construction. No experiment runs a uniform or goal-biased sampler inside the same BiRRT* framework, nor a learned sampler without bidirectional search. Consequently, the 99-100% success rates and large sample reductions may reflect bidirectional expansion rather than the Perceiver-IO sampling distribution. The Gaussian representational question is secondary: without an ablation, the data cannot distinguish capacity failure from measurement confound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SIL-RRT*, a learning-based motion planning algorithm that extends RRT* by using a Perceiver-IO-based neural network to predict a multivariate Gaussian sampling distribution conditioned on point-cloud obstacle representations, the goal, and the last five tree nodes. The network is trained initially by maximum-likelihood on paths collected by RRT*, then fine-tuned with weighted self-imitation learning (WSIL) that weights training trajectories by the discrepancy between actual and estimated path length. The authors evaluate SIL-RRT* against RRT* and MPNetNR in 2D, rigid-body, and 3D environments, and also report results on more complex scenes including a 5-DoF snake robot, claiming large reductions in the number of samples and computation time while producing near-optimal paths.","tokens_in":112,"tokens_out":3884,"duration_ms":68231,"significance":"If the central claim were established, SIL-RRT* would be a useful contribution: it demonstrates that a Transformer-based sampler can be trained from self-generated data without expert demonstrations, and it reports generalization to held-out workspaces. The paper has concrete strengths: it tests on unseen scenarios, evaluates across several environment types, runs the neural network on CPU during planning, and avoids post-processing tricks. However, the evaluation as designed cannot separate the contribution of the learned sampler from other algorithmic modifications (notably bidirectional search) and is weakened by asymmetric sample caps, very small numbers of trials, missing baselines in the complex scenarios, and an uncontrolled comparison with MPNetNR. The significance of the empirical claims is therefore currently conditional on additional controlled experiments.","major_comments":[{"comment":"The central claim in Section 5.2 is that SIL-RRT* achieves reduced sample requirements because of its learned sampling distribution. However, Section 4.1 states that SIL-RRT* integrates the BiRRT* algorithm, which alternates forward and backward tree expansions, while the RRT* baseline does not use bidirectional search. No ablation isolates the learned sampler's contribution: there is no experiment running a uniform or goal-biased sampler inside the same BiRRT* framework, nor a learned sampler without BiRRT*. The observed gains in Table 1 could therefore reflect the bidirectional search or other implementation details rather than the Perceiver-IO sampling distribution.","section":"Section 4.1"},{"comment":"The sample caps are asymmetric in the 3D environment: RRT* is capped at 400 samples while SIL-RRT* and MPNetNR are capped at 200. With this setup, RRT*'s average samples generated (362) is close to its cap and its success rate (17.20%) is constrained by that cap, so the claim of 'fewer samples' and the success-rate comparison are partly by construction. The 3D numbers should be re-run with a common cap for all algorithms, or the comparison should report samples to first feasible solution under an identical termination condition.","section":"Section 5.2, Table 1"},{"comment":"The number of trials is very small: Table 1 reports three trials per environment and Table 2 reports four trials, yet no statistical significance tests or confidence intervals are provided. Given that the standard deviations are often comparable to or larger than the reported differences (e.g., 2D average samples 22.53 +/- 18.05 versus 23.93 +/- 19.09), the claimed advantages of SIL-RRT* over the alternatives are not statistically supported. The authors should report more trials and appropriate significance testing or effect-size measures.","section":"Section 5.2, Tables 1 and 2"},{"comment":"The complex and snake scenarios contain no baseline algorithms; Table 2 compares only SIL-RRT* without and with WSIL. Consequently, the paper's claim that SIL-RRT* 'scales to more complex environments' and handles high-dimensional problems is not supported by any comparison with RRT* or another non-learning planner in these settings. At minimum, RRT* with the same sample cap should be run in the complex 2D, complex rigid-body, complex 3D, and snake environments.","section":"Section 5.2, Table 2"},{"comment":"The MPNetNR comparison is not well controlled. MPNetNR was trained on the authors' dataset for only 5000 iterations, which the paper itself notes is smaller than the dataset used in the original MPNet work, and MPNetNR's point clouds are sampled from obstacle interiors while SIL-RRT* uses surface point clouds. The poor MPNetNR results are therefore attributable at least in part to these training and representation differences, not necessarily to algorithmic inferiority. Additionally, MPNetNR's sample count is not reported in Table 1 (shown as a dash), so the sample-efficiency comparison is incomplete.","section":"Section 5.2, Table 1"}],"minor_comments":[{"comment":"The first paragraph of Section 4 and all of Section 3.3 are essentially identical, including the same discussion of imitation learning and the same references; one of the two copies should be removed or replaced with a brief pointer.","section":"Section 3.3 and Section 4"},{"comment":"The text says 'we executed tree trials for each environment'; this should read 'we executed three trials', and later 'we conducted 4 additional trials' should be 'four additional trials' for consistency.","section":"Section 5.2"},{"comment":"In the 3D row, the Average Time for SIL-RRT* with WSIL is reported as '0.17 +/- 0.4 s'; the standard deviation 0.4 is inconsistent with the neighboring value '0.17 +/- 0.03 s' for SIL-RRT* without WSIL and appears to be a typo.","section":"Table 1"},{"comment":"The notation of the sampler loss is unclear: 'L_sampler = - 1/B NX i=1 log ...' lacks a summation symbol and an explicit definition of the index range, making the equation difficult to parse as written.","section":"Equation (3)"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely a legitimate contribution if the authors can provide controlled ablations that isolate the learned sampler from bidirectional search and equalize the evaluation conditions. The current evidence does not rule out the possibility that the reported gains come from BiRRT* or from the asymmetric sample caps. I would encourage the editor to request the additional experiments described in the major comments rather than to reject outright, because the proposed mechanism is plausible and the implementation appears nontrivial."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe stress-test has it right. The headline result—that SIL-RRT* needs far fewer samples than RRT*—is not currently evidence for the learned sampling distribution. The method integrates BiRRT* (forward/backward tree expansion) specifically to fix feasibility failures, but the RRT* baseline is plain RRT* without that machinery. And in 3D the sample cap is 400 for RRT* and 200 for SIL-RRT*, so 'fewer samples' is partly by construction. Until there is an ablation that runs a uniform or goal-biased sampler inside the same BiRRT* framework, the speedups are confounded.\n\nThat said, the paper is not a throwaway. What is actually new is the combination of a Perceiver-IO encoder/decoder as the sampler with a learned path-length estimator used for self-imitation weighting. The architecture is described clearly enough to reimplement, and the WSIL formulation (Equations 4–6) is a sensible way to down-weight poor demonstrations without needing hand-collected expert data. The snake scenario is a genuine high-dimensional test, even though no baselines are reported for it. The authors also state their limitations honestly: the dataset-size dependency and the fixed robot-size/shape assumption are both in Section 6.\n\nThe softer spots are the standard ones. Three trials per environment is too few to make 99% vs 71% claims stick without confidence intervals. MPNetNR is trained on a dataset much smaller than its original paper used, so its poor showing is not a fair comparison. The complex and snake tables have no RRT* or MPNetNR columns at all, making them descriptive rather than comparative. There is also a duplicated block: the imitation-learning background from Section 3.3 reappears verbatim as the start of Section 4. That is an editing slip, not a scientific flaw, but it should have been caught.\n\nThe central argument can be fixed. The method is plausible, and the held-out workspace evaluation is a reasonable design. A serious referee should take it, but the revision needs an ablation that separates the bidirectional search from the learned distribution, more trials with statistics, and fair baseline training. If the ablation still shows a large sample reduction, this becomes a useful result for the learned-sampling subfield. Without it, the current numbers don't support the abstract.\n\nI would send this to peer review with major revisions. A reading group could learn from it as a case study in confounded system evaluation.","headline":"BiRRT* confound makes the headline sample reductions uninterpretable, but the architecture and WSIL idea are worth a serious referee.","tokens_in":10752,"tokens_out":2955,"would_cite":false,"duration_ms":29476,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SIL-RRT* combines RRT* with a transformer trained by self-imitation to predict a Gaussian distribution for the next sample, and the paper reports it solves 2D, rigid-body, 3D, and 5-DOF snake planning with far fewer samples than RRT* or…","keywords":["motion planning","RRT*","sampling distribution learning","self-imitation learning","transformer","Perceiver-IO","point cloud obstacle representation","bidirectional RRT*"],"falsifier":"Run SIL-RRT* with the trained sampler replaced by a uniform sampler while leaving BiRRT*, goal handling, obstacle representation, and sample caps unchanged; if success rates and average sample counts remain at the reported levels, the learned distribution is not what carries the result.","tokens_in":60,"feed_emoji":"🤖","tokens_out":12273,"duration_ms":170541,"temperature":0.7,"pith_summary":"The paper proposes SIL-RRT*, a version of RRT*—the standard sampling-based planner that grows a collision-free tree toward a goal—in which the next sample is not drawn uniformly but from a distribution predicted by a neural network. The network is a Perceiver-IO transformer that reads the goal, a point-cloud description of obstacle surfaces, and the last five nodes of the growing tree, then outputs the mean and standard deviation of a Gaussian over the next sample. Training starts from paths collected by RRT* and is then fine-tuned with weighted self-imitation learning: the algorithm reuses its own successful paths, weighting each by how much shorter it is than a learned estimate of expected path length. The paper claims that across 2D, rigid-body, 3D, and 5-DOF snake environments this reduces the number of samples needed to find feasible solutions by a large margin, raises success rates relative to RRT* and MPNetNR, and gives paths that are comparable or shorter. The reason this matters is that sample efficiency is the main obstacle to applying sampling-based planners in high-dimensional and obstacle-dense problems.","feed_headline":"Trained RRT* sampler solves 3D planning with ~20x fewer samples","feed_subtitle":"SIL-RRT* predicts a Gaussian sampling distribution from obstacle points and its own paths, beating RRT* and MPNetNR.","key_machinery":"The load-bearing object is the Gaussian sampler network: a Perceiver-IO transformer whose encoder projects obstacle-surface point clouds into a fixed-length latent array and whose causal decoder attends to the goal plus the last five tree nodes. Two MLP heads convert the decoded features into the mean and standard deviation of a multivariate normal distribution, and samples are drawn from that distribution to grow the RRT*. Training minimizes the negative log-likelihood of the nodes of feasible paths, with each path weighted by a sigmoid of the gap between its true length and the prediction of a separate estimator network, plus an entropy term to encourage exploration. The same architecture is paired with an integrated bidirectional RRT* search, added to overcome difficulty in generating useful samples near the goal.","core_discovery":"The central claim is that a sampling distribution for RRT* can be learned from data the algorithm itself generates, and that this learned distribution transfers to workspaces the model has not seen. In the reported comparisons, SIL-RRT* uses on average about 4 times fewer samples than RRT* in 2D and rigid-body scenes and about 20 times fewer in the 3D suite, reaches 100 percent success in 3D where RRT* succeeds only 17.2 percent even though RRT* was allowed a 400-sample budget and SIL-RRT* only 200, and produces paths that are shorter or comparable in length. The fine-tuned version with weighted self-imitation learning further shortens path lengths at a modest cost in samples. The paper presents these results as evidence that a transformer-based, self-imitation-trained sampler can replace uniform sampling across different configuration-space dimensions.","pith_inferences":["Beyond the paper's claims, a clean ablation that replaces the learned sampler with a uniform sampler while keeping BiRRT*, goal handling, and sample caps fixed would show whether the learned distribution is actually the cause of the reported speedups; the paper does not report such an ablation.","Beyond the paper's claims, the single Gaussian output head may be the bottleneck in environments with narrow passageways where the useful next-sample distribution is multi-modal; a mixture-density or diffusion head would be a direct test of that limitation.","Beyond the paper's claims, the learned sampler could serve as a proposal distribution inside other sampling-based planners, such as informed or batch-informed tree search, rather than only inside RRT*.","Beyond the paper's claims, the five-node context window limits the sampler to local tree history; environments whose corridors require very long-range dependencies may need a larger context or a recurrent summary."],"forward_implications":["A planner built this way can solve new workspaces that were not seen in training, since the test set used unseen scenarios with more obstacles than the training set.","In the 3D suite, SIL-RRT* solves every task within 200 samples, whereas RRT* solves only 17.2 percent within 400 samples, suggesting learned sampling can extend the reach of sampling-based planning into higher dimensions.","Because weighted self-imitation learning reuses the planner's own successful paths, the method avoids the need for near-optimal expert demonstrations that are difficult to obtain in high-dimensional spaces.","The same architecture and surface-point-cloud representation are reported to work for point-mass, rigid-body, and 5-DOF snake agents, so the approach is not tied to one configuration-space dimension."],"supporting_citations":[{"why":"Defines RRT*, the algorithm being extended and the baseline that must be beaten.","marker":"[19]"},{"why":"Motion Planning Networks, the supervised-learning baseline SIL-RRT* is compared against.","marker":"[27]"},{"why":"Perceiver encoder that projects obstacle point clouds into a fixed-length latent array.","marker":"[17]"},{"why":"Perceiver-IO decoder that produces the sampler's Gaussian output from the latent and node sequence.","marker":"[16]"},{"why":"Bidirectional RRT* variant integrated into SIL-RRT* to handle difficult goal sampling.","marker":"[18]"},{"why":"Self-imitation learning method that weighted self-imitation learning draws on for reusing the agent's own successful paths.","marker":"[26]"},{"why":"Obstacle representation as point clouds sampled from obstacle surfaces.","marker":"[30]"},{"why":"Decision Transformer technique of conditioning on the last five nodes with a causal mask.","marker":"[6]"}],"fun_headline_variants":["RRT* learns its sampling from itself, cuts 3D samples 20x","Self-imitation sampler: 20x fewer samples, 100% success in 3D","SIL-RRT*: AI-guided sampling makes RRT* 20x more efficient","Trained RRT* sampler: 4x fewer samples in 2D, 20x in 3D","No more uniform sampling: RRT* learns to pick better points"],"cache_read_input_tokens":12928,"weakest_assumption_plain":"The whole approach depends on the assumption that one bell-shaped probability distribution, computed from the goal, the obstacle point cloud, and the last five tree nodes, is enough to steer RRT* efficiently; if that assumption fails, the reported speedups could come from the built-in bidirectional search rather than from the learned sampler.","fun_headline_variants_meta":{"raw":{"variants":["RRT* learns its sampling from itself, cuts 3D samples 20x","Self-imitation sampler: 20x fewer samples, 100% success in 3D","SIL-RRT*: AI-guided sampling makes RRT* 20x more efficient","Trained RRT* sampler: 4x fewer samples in 2D, 20x in 3D","No more uniform sampling: RRT* learns to pick better points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000891,"raw_usage":{"total_tokens":3787,"prompt_tokens":830,"completion_tokens":2957,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":446,"completion_tokens_details":{"reasoning_tokens":2840}},"tokens_in":446,"tokens_out":2957,"duration_ms":19984,"temperature":1.0,"reasoning_tokens":2840,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:16:02.651664+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SIL-RRT* with the trained sampler replaced by a uniform sampler while leaving BiRRT*, goal handling, obstacle representation, and sample caps unchanged; if success rates and average sample counts remain at the reported levels, the learned distribution is not what carries the result.","supporting_citations":[{"cited_title":"Perceiver: General perception with iterative attention","cited_arxiv_id":null,"evidence_quote":"Perceiver encoder that projects obstacle point clouds into a fixed-length latent array."},{"cited_title":"Optimal bidirectional rapidly- exploring random trees","cited_arxiv_id":null,"evidence_quote":"Bidirectional RRT* variant integrated into SIL-RRT* to handle difficult goal sampling."},{"cited_title":"Sampling-based algorithms for op- timal motion planning","cited_arxiv_id":null,"evidence_quote":"Defines RRT*, the algorithm being extended and the baseline that must be beaten."},{"cited_title":"Motion planning networks","cited_arxiv_id":null,"evidence_quote":"Motion Planning Networks, the supervised-learning baseline SIL-RRT* is compared against."},{"cited_title":"Self-imitation learning","cited_arxiv_id":null,"evidence_quote":"Self-imitation learning method that weighted self-imitation learning draws on for reusing the agent's own successful paths."},{"cited_title":"Learning obstacle representa- tions for neural motion planning","cited_arxiv_id":null,"evidence_quote":"Obstacle representation as point clouds sampled from obstacle surfaces."},{"cited_title":"Deci- sion transformer: Reinforcement learning via sequence modeling","cited_arxiv_id":null,"evidence_quote":"Decision Transformer technique of conditioning on the last five nodes with a causal mask."}],"review_version":1}