{"id":"5d911b28-23b8-4e27-87c3-3aa92c073a43","arxiv_id":"2601.22965","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SIDP trains a diffusion policy for visual navigation by reward-weighting its own sampled trajectories, improving success rate and cutting inference latency.","lead":"This paper introduces SIDP, a diffusion-policy navigation system that trains itself by imitating its own high-reward trajectories, reaching higher success rates than the prior NavDP baseline. It reports a 2.5× inference speedup on an embedded GPU and qualitative real-world demos on two robots, which makes it relevant to real-time visual navigation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported performance numbers are internally inconsistent: the abstract's ~10-point Commercial margin contradicts Table II's 5.79-point difference, and Table III's NavDP SR (0.549) conflicts with Table II's NavDP mSR (71.25%).","rationale":"The reader's weakest assumption (ESDF reward fidelity for sim-to-real transfer) is a valid concern but does not directly invalidate the primary quantitative claims, which are made in simulation. A more immediate and concrete problem is that the reported headline numbers are internally inconsistent. The abstract's '~10 points on InternScene-Commercial' is contradicted by Table II, and Table III's SR values for the same methods disagree with Table II unless a different evaluation protocol is being used. This makes the central claims of SOTA performance and the 2.5× speedup difficult to verify, and could indicate either a reporting error or an unstated evaluation setting. The paper should be accepted only after these inconsistencies are resolved. Hence, the verdict remains CONDITIONAL, matching the reader's overall assessment, though the specific focus differs.","tokens_in":11342,"tokens_out":9334,"duration_ms":104886,"concrete_test":"Two checks: (1) Recompute the Commercial and mSR margins from Table II and verify which setting corresponds to the abstract's '~10 points' claim. (2) Run the exact NavDP inference pipeline (including its auxiliary critic and sampling protocol) on the same evaluation episodes and hardware used to produce Table III; confirm whether NavDP's SR is ~0.549 and whether SIDP's 110ms/0.674 numbers are measured on that same protocol. If Table III uses the custom one-shot benchmark from §IV-A1, the paper must explicitly say so and align the metric scaling (percentage vs fraction).","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central quantitative claims are undermined by unexplained inconsistencies. The abstract states SIDP surpasses NavDP by approximately 10 points on the InternScene-Commercial setting, but Table II shows Commercial SR: NavDP 57.38, SIDP 63.17 — a 5.79-point margin, not ~10. The ~10-point figure matches the mSR difference (73.22→81.19, i.e., 9.94 points), which is an overall average, not the Commercial subset. More importantly, Table III reports NavDP SR=0.549 and SIDP SR=0.670 for 10-step DDPM, whereas Table II lists NavDP mSR=71.25% and SIDP mSR=81.19%. If Table III uses the same evaluation protocol, these numbers are irreconcilable; if it uses a different benchmark (e.g., the custom one-shot setting described in §IV-A1), the paper never states this. Without clarification, the 2.5× speedup claim (110ms vs 273ms) is not tied to a clearly specified success-rate context. These inconsistencies do not disprove the method, but they erode confidence in the headline results, which are the paper's primary evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SIDP, a self-imitating diffusion-policy training scheme for visual navigation. Instead of imitating fixed expert demonstrations, SIDP samples multiple candidate trajectories from the current policy, scores them with a hand-designed reward based on ESDF collision checks and goal progress, keeps the top-k by reward, and updates the policy with a reward-weighted denoising loss. Two auxiliary mechanisms are added: goal-agnostic exploration and a reward-driven curriculum. Experiments on the InternVLA-N1 S1 benchmark report state-of-the-art SR and mSPL over NavDP and other baselines, and an inference-latency comparison on Jetson Orin Nano reports up to a 2.5x speedup. Real-world deployments on two robots are shown qualitatively.","tokens_in":11743,"tokens_out":10583,"duration_ms":103370,"significance":"If the results hold, the work makes a useful contribution: it offers a stable, BPTT-free way to improve a diffusion policy from its own samples, and it removes the auxiliary trajectory selector used by prior diffusion planners, which matters for edge deployment. The derivation in Eqs. (3)-(8) is a standard REPS/advantage-weighted regression scheme and is internally consistent; the self-referential element is not circular because the target is defined by environment reward. The paper also includes a clear algorithm and several ablation experiments. However, the empirical evidence is currently weakened by reporting inconsistencies, missing statistical detail, and an under-specified training protocol, so the central quantitative claims need revision before the contribution can be assessed.","major_comments":[{"comment":"The abstract's headline claim that SIDP surpasses NavDP by 'approximately 10 points on the InternScene-Commercial setting' is not supported by Table II. The Commercial SR margin is 5.79 points (NavDP 57.38 vs SIDP 63.17). The 9.94-point figure corresponds to the mSR column (71.25 vs 81.19). The sentence in §IV-C1 saying 'In the Commercial and Home scenarios ... margins of 9.94% and 5.79% in Success Rate, respectively' is also inconsistent with the table: Home SR is NavDP 93.37 vs SIDP 94.36, a 0.99-point difference. These numbers must be corrected and disambiguated.","section":"Abstract and §IV-C1, Table II"},{"comment":"Table III reports SR values (NavDP 0.549, SIDP 0.670 for 10-step DDPM) that cannot be reconciled with Table II (mSR 71.25% and 81.19%). If Table III was run on the custom one-shot benchmark described in §IV-A1, this must be stated explicitly in the table caption and text; if it is the same InternVLA-N1 S1 protocol, the numbers conflict. In addition, no standard deviations or number of seeds are provided for any main-table result, so the reader cannot tell whether the reported margins (e.g., Home SR +0.99, or the DDIM 5-step SR 0.674 vs 10-step 0.670) are meaningful. Please provide seeds/error bars and state the evaluation protocol for each table.","section":"§IV-A1 and Table III"},{"comment":"The training procedure is described as on-policy self-imitation (Algorithm 1 samples states from 'environment E' and generates actions from the current policy), but the implementation in §IV-B1 samples initial poses from expert trajectories in InternData-N1 and evaluates trajectories offline with ESDF queries, without rolling out the policy. The objective in Eq. (3) is an expectation over the policy's state distribution D, so optimizing with states drawn from a fixed expert dataset is a distribution mismatch. This undermines the theoretical grounding of the self-imitation loop and the robustness claim. Please either present the method as an offline/off-policy variant, resample states from policy rollouts (as in DAgger), or justify why expert-state resampling suffices.","section":"§III-A, §IV-B1, Algorithm 1"},{"comment":"The headline 2.5x speedup (110 ms vs 273 ms) is obtained by comparing SIDP with 5-step DDIM to NavDP with 10-step DDPM. The table also shows SIDP with 10-step DDPM already runs in 132 ms, so part of the gain comes from removing the selector/critic and part from scheduler/step reduction. To substantiate the claim that self-imitation enables the efficiency gain, the authors should run NavDP with DDIM and with fewer steps, or argue quantitatively why this is not a valid control. Without that, the 2.5x figure conflates algorithm and scheduler choices.","section":"§IV-C2, Table III"},{"comment":"The real-world experiments are purely qualitative (Fig. 7 time-lapse images). The abstract states 'real-world experiments confirming its effectiveness across multiple robotic platforms.' Given that the perception stack in the real world uses BANet-estimated depth while training uses clean simulation/ESDF, the sim-to-real claim needs quantitative support: at minimum, number of trials, success rate, collision count, and goal error on each platform. Without this, the claim of robust sim-to-real transfer is not supported.","section":"§IV-E and Abstract"}],"minor_comments":[{"comment":"Several hyperparameters are not specified: the candidate count N, truncation k, and the curriculum thresholds τ_max and τ_range. Please report these values to make the method reproducible.","section":"§IV-B2 and §III-C2"},{"comment":"The figures do not label their y-axes or define the plotted metric in the current text; e.g., Fig. 4's caption says 'ablation study...' without stating whether the vertical axis is SR or SPL. Please add axis labels and metric definitions.","section":"Figures 3-6"},{"comment":"The notation in Eq. (8) is garbled: 'a(i),tdn' should be something like \\hat{a}^{(i)}_{t_{dn}}. Also, the reduction from weighted log-likelihood to the denoising loss should cite the standard DDPM variational bound explicitly, since the per-trajectory weights interact with the per-timestep loss in a non-obvious way.","section":"Eq. (8)"},{"comment":"The claim that SIDP 'outperforms existing methods in both SR and SPL' should be qualified. In the Home category, SIDP's SPL is lower than NavDP's (89.86 vs 91.44), even though the mean SPL is higher. The text should acknowledge this category-level exception.","section":"§IV-C1 and Table II"},{"comment":"The paper does not mention code or data release. Given that the training scenario and reward evaluation are custom, releasing code/data or providing detailed implementation details would substantially improve verifiability.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sensible and the derivation is internally consistent, so I see this as a fixable paper. The main obstacles are the inconsistent headline numbers, the unclear protocol for Table III, the offline-vs-online state distribution mismatch, and the qualitative-only real-world evidence. I would like the authors to correct the reporting, add seeds/error bars, and clarify the training/evaluation protocol. If the numbers do not survive reruns with error bars, the contribution would be much weaker, but I do not think rejection is warranted at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Core idea is worth a look: reward-guided self-imitation for diffusion policies, applied to visual navigation. The derivation in Eqs. 3-8 is standard REPS-style importance weighting with top-k truncation and is internally consistent. What's new is the packaging: self-sampled trajectories as training distribution, no backprop through denoising, and then dropping the auxiliary trajectory selector for inference with fewer DDIM steps. If the empirical claims hold, that's a real engineering win for embedded robots.\n\nWhat the paper does well: related work is appropriate, method is readable, ablations are well designed, comparing linear weighting and sweeping temperature are the right checks. Curriculum and goal-agnostic exploration are plausible additions, though their individual contributions look modest.\n\nSoft spots are mostly in reporting. The stress-test is on target: abstract says \"approximately 10 points\" on Commercial, but Table II shows 5.79 points on Commercial SR; the 9.94-point margin is overall mSR, not Commercial. Section IV-C repeats the mismatch, attributing 9.94% to Commercial and 5.79% to Home, which doesn't match the table either. Table III reports SR as decimals (0.549, 0.670) that match neither the percentages in Table II nor any stated benchmark - the paper never says whether that table uses the main closed-loop benchmark or the custom one-shot ablation. That's load-bearing, because the 2.5x speedup is tied to that table.\n\nAlso, no error bars or seeds are reported, no code released, and real-world results are qualitative only. Those are fixable, but they limit reproducibility. The central argument, that self-imitation concentrates the trajectory distribution and enables efficient inference, holds up conceptually. The reward is external to the policy, so the circularity concern is minor.\n\nThis paper is for people building diffusion-based navigation policies for edge hardware. It deserves a serious referee: the idea is sound and the payoff is concrete. I would not cite it yet, but I'd engage with it. Ask the authors to reconcile the tables, clarify the Table III benchmark, and provide variance estimates.\n\nNet: send to peer review with a request for major revisions on empirical reporting.","headline":"Reward-weighted self-imitation for diffusion navigation is a sound combination, but the reported numbers have internal inconsistencies that need fixing before the headline claims are credible.","tokens_in":721,"tokens_out":1653,"would_cite":false,"duration_ms":38521,"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":"The paper proposes that a diffusion policy for visual navigation can learn better paths by selectively imitating its own high-reward sampled trajectories, and that this self-imitation removes the need for expensive generate-then-filter infe","keywords":["visual navigation","diffusion policy","self-imitation learning","reward-guided learning","trajectory planning","curriculum learning","edge deployment","success weighted by path length"],"falsifier":"Run SIDP on the benchmark with the reward components stripped down, such as only the collision penalty or only the step cost; if SR and SPL collapse to NavDP levels, the reported gains come from reward design rather than the self-imitation mechanism itself. Alternatively, a quantitative real-world trial with 100 or more episodes comparing SIDP's 5-step DDIM no-selector inference against NavDP would settle whether the 2.5x speedup and success-rate margin survive outside simulation.","tokens_in":11284,"feed_emoji":"🤖","tokens_out":4915,"duration_ms":47715,"temperature":0.7,"pith_summary":"The paper tries to establish that a diffusion policy for visual navigation does not have to imitate external expert demonstrations and does not need a generate-then-filter inference pipeline. Instead, the policy samples candidate trajectories from itself, keeps the ones that score well under a safety-and-efficiency reward, and trains on those with a reward-weighted loss, concentrating the trajectory distribution around high-quality modes. If true, diffusion-based navigation can run end-to-end on embedded hardware because the auxiliary critic and dense sampling become unnecessary. The authors report higher Success Rate and Success weighted by Path Length than prior methods on the InternVLA-N1 S1 benchmark, a 2.5x latency reduction on Jetson Orin Nano, and qualitative real-world deployments on two robots.","feed_headline":"A robot navigator that imitates its own best paths runs 2.5x faster","feed_subtitle":"Self-generated high-reward trajectories replace expert demos, cutting edge inference from 273 ms to 110 ms.","key_machinery":"The central mechanism is the reward-weighted denoising loss L_SIDP, built from reward-guided self-imitation. The policy acts as its own proposal distribution: candidate trajectories are drawn from the current policy, scored by a hand-designed reward, the top-k are kept and converted into softmax importance weights, and those weights multiply the standard diffusion denoising error. This converts the intractable goal of matching an optimal trajectory distribution into a stable iterative imitation objective, avoiding backpropagation through denoising steps while still shaping the policy toward high-reward trajectories. Two auxiliary mechanisms support it: goal-agnostic exploration expands the t","core_discovery":"On the paper's own terms, the central discovery is that a diffusion policy's own sampled trajectories, reweighted by reward, are a better training target than expert demonstrations. Sampling N candidates from the current policy, truncating to the top-k by a reward combining collision avoidance, path efficiency, progress, and final docking, then computing a softmax-weighted denoising loss, approximates minimizing KL divergence to an optimal trajectory distribution. Iterating this update makes the policy output concentrated and consistent, so inference can use deterministic DDIM sampling with as few as five denoising steps and no selector. The reported results include 81.19% mean SR versus 71.","pith_inferences":["The biggest unspoken transfer risk is the reward itself: it is computed from ESDF values in synthetic scenes, while the real-world robots use a learned depth estimator, so the simulation reward may not capture real-world safety; a quantitative real-world benchmark with the same metrics would test this.","A deliberately harder benchmark with scarce or suboptimal expert demonstrations would test the claim that SIDP degrades less than imitation learning when expert data quality drops.","The curriculum thresholds tau_max and tau_range are fixed hyperparameters; an adaptive version is a natural extension, since the paper's own ablation shows they control training stability.","The efficiency gain could be compounded by combining self-imitation with explicit goal-conditioning or memory, since goal-agnostic exploration is presented as a regularizer rather than a replacement for point-goal conditioning."],"forward_implications":["Diffusion navigation policies can be deployed on resource-constrained robots without an auxiliary trajectory selector, because the learned distribution is already concentrated.","Fewer denoising steps such as five-step DDIM become feasible without degrading success, reducing inference latency from 273 ms to 110 ms on an edge GPU.","Training no longer depends on the coverage and quality of expert demonstrations, since self-generated high-reward trajectories can improve the policy iteratively.","The same reward-weighted self-imitation objective applies to any diffusion policy that can be scored by a reward, suggesting reuse for other sequential decision tasks."],"fun_headline_variants":["Robot learns from its own best paths to navigate 2.5x faster","Self-imitated diffusion policy cuts navigation inference to 110ms","Diffusion policy imitates its own top trajectories for robust navigation","Reward-guided self-imitation makes robot navigation faster and sturdier","Autonomous navigation speeds up via self-imitation of high-reward paths"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the hand-designed reward computed from ESDF in the simulated interactive environment is a faithful proxy for real-world navigation quality, so self-imitated high-reward trajectories in synthetic scenes remain high-quality on physical robots with different depth perception.","fun_headline_variants_meta":{"raw":{"variants":["Robot learns from its own best paths to navigate 2.5x faster","Self-imitated diffusion policy cuts navigation inference to 110ms","Diffusion policy imitates its own top trajectories for robust navigation","Reward-guided self-imitation makes robot navigation faster and sturdier","Autonomous navigation speeds up via self-imitation of high-reward paths"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1140,"prompt_tokens":756,"completion_tokens":384,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":291}},"tokens_in":500,"tokens_out":384,"duration_ms":4421,"temperature":1.0,"reasoning_tokens":291,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T06:16:51.695380+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SIDP on the benchmark with the reward components stripped down, such as only the collision penalty or only the step cost; if SR and SPL collapse to NavDP levels, the reported gains come from reward design rather than the self-imitation mechanism itself. Alternatively, a quantitative real-world trial with 100 or more episodes comparing SIDP's 5-step DDIM no-selector inference against NavDP would settle whether the 2.5x speedup and success-rate margin survive outside simulation.","supporting_citations":[],"review_version":1}