{"id":"8ea2a418-4809-44d7-a87e-3a96af8e71b2","arxiv_id":"2504.19322","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A learned perceptive forward dynamics model, trained on simulation and real data, improves pose prediction by 41% and navigation success in rough sim terrain when used inside an MPPI planner.","lead":"This paper trains a neural network to predict a legged robot's future motion and collision risk from its current surroundings and sensor history, then plugs that network into a sampling-based planner. The result is safer rough-terrain navigation with a much simpler reward function that needs less hand-tuning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Navigation-success gain is measured against a Kim et al. baseline trained only on 2D data and tested in 3D; the 27% claim lacks a like-for-like comparison.","rationale":"The reader's stated weakest assumption is the single-input height scan, which is a real limitation but is partly mitigated by the 7 Hz receding-horizon planner: a fresh height scan arrives every 0.14 s, so a static 5 s prediction inside one MPPI iteration is not the final word on safety. The planning-baseline mismatch, by contrast, is not mitigated by the architecture and directly inflates the headline success-rate number. The paper itself documents the unequal training split, making the concern concrete and checkable rather than speculative. I therefore select it as the most load-bearing issue. The reader's CONDITIONAL verdict remains appropriate; the concern strengthens the need for clarification but does not overturn the paper's core contribution.","tokens_in":19631,"tokens_out":4827,"duration_ms":53052,"concrete_test":"Retrain the Kim et al. baseline on exactly the same combined 2D/3D (and 2D-3D) synthetic dataset and training schedule as the proposed FDM, then rerun the Table III planning evaluation in the 3D environment with the same MPPI tuning protocol and report per-seed success rates. If the retrained baseline still shows roughly 25-percentage-point lower success, the architecture comparison stands; if it approaches Ours, the headline success-rate gain is largely a training-data artifact. If the 3D-trained baseline is truly too unreliable to plan, quantify that failure (success rate, collision counts) instead of substituting a 2D-trained model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central downstream claim is the 27% higher navigation success in rough environments. In Table III the Kim et al. baseline reaches 78.33% (2D) / 48.75% (3D), while Ours reaches 88.33% / 73.75%. Section VI-E states that this baseline 'was trained solely in a 2D environment, as training in a more complex 3D environment made the collision predictions, as expected, unreliable.' That means the 3D comparison is a domain-shift test: the baseline sees 3D geometry through a 2D-LiDAR model trained only on 2D layouts. The position-estimation experiment (Sec. VI-B) carefully retrains Kim et al. on the same data as Ours and is therefore fair, but this fairness is not carried into the planning experiment, where the headline success-rate gain is produced. The 3D margin could be caused by the baseline's training-distribution mismatch rather than by the proposed perceptive FDM. Table III also lacks repeated-run statistics or confidence intervals, so the reported differences are not yet shown to be significant. The concern is not that the method is weak; it is that the strongest claim is currently supported by an unequal comparator.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a learned perceptive Forward Dynamics Model (FDM) for quadrupedal robots, predicting future SE2 poses and failure risk over a 5-second horizon from proprioceptive histories, velocity actions, and a current height scan. The model is pre-trained on large-scale simulated data, optionally fine-tuned with real-world data, and integrated into an MPPI planner using a two-term reward composed of pose error and predicted failure risk. Experiments on ANYmal, Barry, and ANYmal-on-Wheels report improved position prediction over the Kim et al. baseline and constant-velocity assumption, higher planning success in 2D and 3D simulation, and qualitative real-world transfers. The headline quantitative claims are a 41.28% average reduction in final-step position error and a 27% higher navigation success rate in rough simulation environments.","tokens_in":19890,"tokens_out":5337,"duration_ms":55132,"significance":"If the claims hold, this is a valuable contribution to legged navigation: it demonstrates that a learned perceptive dynamics model can replace hand-designed traversability costs in sampling-based planning, and the authors provide public code and models, modality ablations in Appendix D, platform-aware experiments across three robots, and a real-world fine-tuning study. The 41% position-error improvement in Table II is well supported because the Kim et al. baseline is retrained on the same data for that experiment. However, the downstream 27% success-rate claim is currently supported by an unequal planning comparison and lacks statistical confidence, so the headline result is not yet fully established.","major_comments":[{"comment":"The headline downstream claim—27% higher navigation success in rough environments—rests on an unfair comparison. The text explicitly states that the Kim et al. baseline 'was trained solely in a 2D environment,' while the proposed FDM uses full 3D data, so the 3D success-rate gap (73.75% vs. 48.75%) could be caused by training-distribution mismatch rather than by the proposed model. To support the claim, the authors should retrain the baseline on the same 3D data, as was done for the dynamics comparison in §VI-B, or clearly reframe the experiment as a domain-shift test and add a like-for-like baseline.","section":"§VI-E and Table III"},{"comment":"The planning results report success rates without any number of trials, starts, seeds, or confidence intervals. Given that the 2D success rates are 88.33%, 78.33%, and 82.50%, the 10-percentage-point difference between the proposed method and Kim et al. may not be significant; the same concern applies to the 3D numbers. The authors should report the number of runs per condition and provide repeated-run statistics, such as bootstrap confidence intervals or per-seed standard deviations.","section":"Table III"},{"comment":"The height scan ht is encoded once by the CNN and only initializes the hidden state of the forward prediction GRU; no observation is re-encoded during the 10-step, 5-second rollout. The model therefore assumes that the static initial height scan contains all geometric information needed for the entire prediction horizon, which is questionable in cluttered or occluded rough terrain where obstacles become visible only as the robot moves. This assumption should be explicitly acknowledged as a limitation and, ideally, tested with an ablation that re-injects observations at each step or with scenarios where later-visible obstacles are present.","section":"§V-B and Fig. 2"},{"comment":"The abstract describes the approach as a 'heuristic-free solution' and the paper claims that the simplified cost formulation 'eliminates the need for extensive cost-tuning to ensure safety,' but Appendix H states that MPPI parameters such as noise magnitude, time correlation, reward scaling, and sampling space 'require carefully tuned parameters' in complex scenarios. This tension should be resolved by either softening the heuristic-free claim or quantifying the remaining tuning effort, for example with a sensitivity analysis of the MPPI hyperparameters.","section":"§V-D and Appendix H"}],"minor_comments":[{"comment":"The phrase '27% higher navigation success rate' is ambiguous: in Table III the 3D gap is 25 percentage points (73.75% vs. 48.75%), which is not 27% higher in relative terms; the authors should specify whether they report relative improvement, percentage points, or an average over environments, and give the exact arithmetic.","section":"Abstract and Table III"},{"comment":"The caption should define what counts as a successful run and how many runs per environment were executed, since success rate without a trial count is not interpretable.","section":"Table III"},{"comment":"The y-axis label 'Position Delta (m)' with the text 'linear logarithmic' is confusing; the authors should state explicitly which scale is used and how the 95% quantile is displayed.","section":"Fig. 5 and Fig. 13"},{"comment":"Typo: 'Synethic Data Generation' should be 'Synthetic Data Generation.'","section":"§V-A"},{"comment":"Typo: 'Data-Colletion' should be 'Data Collection.'","section":"Fig. 2"},{"comment":"The term 'neighboring paths' is used without a precise definition; the authors should define q and the neighborhood relation explicitly, since Appendix F refers to a 'neighboring filter parameter' that is not formally introduced.","section":"Eq. (15)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid systems contribution with convincing dynamics experiments and ablations, but the headline success-rate claim needs a like-for-like planning baseline and statistical support before it can be accepted as stated. The 41% position-error claim appears well supported; the 27% navigation-success claim does not."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a well-executed incremental step beyond Kim et al.'s learned FDM for quadruped navigation. The height-scan perception and real-world fine-tuning are genuine additions for rough terrain, and the position-estimation evaluation is run fairly — the authors retrain Kim's model on the same data. The 41% position-error reduction therefore carries real weight.\n\nThe planning comparison is where I'd push back. The headline 27% success-rate gain in 3D comes from benchmarking against a Kim et al. baseline that was trained only in 2D, because training it in 3D made its collision predictions unreliable. That turns the 3D comparison into a domain-shift test rather than an equal one. The stress-test note is right: the improvement could come from the baseline's training-distribution mismatch, not from your FDM inherently being better. The 2D comparison is more competitive, but the advertised claim is the 3D one. Table III also reports no error bars or repeated runs, so we cannot tell if the differences are significant.\n\nThe abstract overstates the \"heuristic-free\" benefit. Section VII and Appendix H both admit that MPPI action-distribution and reward parameters still need tuning; the correct phrasing is \"reduced tuning,\" not eliminated. That is a minor framing issue, not a fatal one.\n\nOn the model side, the height scan is injected only once at the start of the prediction horizon, and the GRU decodes future states from actions alone. Over a 5-second horizon in cluttered terrain, that is a real simplifying assumption. The qualitative results look fine, but I'd like to see the authors test or at least explicitly discuss cases where previously occluded geometry becomes relevant mid-horizon.\n\nCredit where due: the ablations in Appendix D cleanly show that proprioceptive history and the height scan both matter, and the real-world fine-tuning results demonstrate sim-to-real transfer. Code and models are public, which raises the reproducibility bar.\n\nWho is this for? Anyone working on learned dynamics or sampling-based navigation for legged robots. It deserves a serious referee: the core idea is sound, the position-estimation evidence is credible, and the planning comparison is fixable by retraining the baseline on the same 3D data and adding repeated-run statistics. I'd recommend conditional acceptance with those requests.","headline":"Solid incremental extension of Kim et al.'s learned FDM, but the headline 27% success-rate gain rests on a 2D-trained baseline and no variance estimates.","tokens_in":20464,"tokens_out":2610,"would_cite":true,"duration_ms":23202,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a learned perceptive forward dynamics model can replace hand-tuned traversability costs for legged-robot navigation, predicting five seconds of future poses and failure risk from a height scan and proprioceptive…","keywords":["forward dynamics model","quadrupedal navigation","MPPI planning","terrain traversability","sim-to-real transfer","perceptive prediction","failure risk prediction","rough terrain"],"falsifier":"Run the trained FDM through a course where a low obstacle is hidden behind a ridge for the first two prediction steps and only becomes visible from the robot's onboard camera later; if the predicted poses keep passing through the obstacle with low risk, the single-scan assumption fails. A quantitative version would measure rollout error against a variant that re-encodes the height scan at every step.","tokens_in":19400,"feed_emoji":"🤖","tokens_out":8186,"duration_ms":81323,"temperature":0.7,"pith_summary":"Rough-terrain navigation usually separates traversability assessment from motion planning and requires hand-tuned cost functions. This paper tries to collapse both into one learned forward dynamics model: given a height scan, a history of poses and proprioceptive readings, and a sequence of velocity commands, the model predicts where the robot will be and how likely it is to fail over the next five seconds. The authors claim that this perceptive prediction makes a sampling-based MPPI planner work with only a goal-distance reward and a risk penalty, with no environment-specific cost tuning, and that training on simulated plus real data transfers to a real quadruped. Their reported numbers on ANYmal are a 41% average reduction in final position error over competitive baselines and a 27% higher navigation success rate in rough simulated environments.","feed_headline":"Learned model cuts robot position error by 41 percent","feed_subtitle":"Trained on simulation and real data, it lets the ANYmal quadruped navigate rough terrain with only a two-term reward.","key_machinery":"The load-bearing object is the perceptive forward dynamics model itself, a compact recurrent network of about 1.16 million parameters. A GRU encoder compresses the past-state and proprioceptive history while a small CNN compresses the height scan; their embeddings initialize the hidden state of a second GRU that rolls the action sequence forward. Two heads decode, at each step, a velocity correction $\\Delta\\tilde{a}$ and a failure probability $\\tilde{r}$; the velocity corrections are integrated into poses, and a \"stop loss\" keeps predicted poses constant once a failure is predicted. On the planning side, the central mechanism is the replacement of hand-designed traversability costs by the model's implicit risk output, so MPPI's reward is $R = \\lambda_{\\text{pose}} R_{\\text{pose}} + \\lambda_{\\text{risk}} R_{\\text{risk}}$ with a terminal goal-distance term and a cumulative neighboring-path risk penalty.","core_discovery":"The paper's central claim is that a learned perceptive forward dynamics model can stand in for both a dynamics simulator and a traversability classifier in local planning. At time $t$, the model receives the current height scan $h_t$, a history of $n$ past states and proprioceptive measurements, and a planned action sequence; a recurrent network predicts the future poses $\\tilde{p}_{t+1},\\dots,\\tilde{p}_{t+n}$ and failure risks $\\tilde{r}_{t+1},\\dots,\\tilde{r}_{t+n}$. The pose prediction is done in residual form: the network outputs velocity corrections that are integrated with a constant-velocity model into $\\mathrm{SE}(2)$ poses, and the risk head is trained with binary cross-entropy. The authors integrate this model into an MPPI planner whose reward is just a terminal goal-distance term and a risk penalty. They report final-step position errors of 0.28 m in 2D, 0.30 m in 2D-3D and 0.28 m in 3D environments versus 0.37--0.45 m for the perceptive baseline, a 41.28% average error reduction, and planning success rates of 88.33% (2D) and 73.75% (3D) versus 78.33% and 48.75% for the baseline FDM planner.","pith_inferences":["Because the height scan enters only once, the model treats the world as static for five seconds; an obvious extension is to re-feed perception at every prediction step, which would also let the planner handle obstacles revealed as the robot moves.","The failure-risk head could double as an uncertainty or curiosity signal for data collection: a planner could seek out trajectories where predicted risk is high but not catastrophic, actively closing the model's blind spots.","The same residual velocity-integration design could be embedded as a differentiable world model for model-based policy optimization or for refining physics simulators, consequences the paper mentions only as future work.","A direct test of the framework's limits would be navigation among moving agents, since the current observations and failure labels assume static geometry; the paper states it has not tested this regime."],"forward_implications":["If the claim is right, sampling-based planners for legged robots can be deployed in new rough environments by retraining or fine-tuning one FDM instead of tuning traversability cost maps.","A failure-risk estimate learned from data, rather than from a geometric heuristic, should transfer across obstacle shapes such as walls, stairs, and ramps without per-obstacle thresholds.","The same architecture and training recipe can be reused for different embodiments: the paper demonstrates platform-aware predictions for a quadruped, a wheeled-legged platform, and another quadruped under the same action sequence.","Real-world fine-tuning becomes a small-data correction step: adding safe human-guided trajectories from snow, forest, or pavement cuts position error by about 30% in those domains.","Because the planner still samples and scores many trajectories, planning behavior can be adjusted at deployment by changing reward weights, noise, and command ranges without retraining the model."],"supporting_citations":[{"why":"Provides the first learned FDM for quadruped navigation and the perceptive baseline against which the 41% position-error improvement is measured.","marker":"[5]"},{"why":"Supplies the residual velocity formulation the FDM uses to avoid predicting absolute poses directly.","marker":"[12]"},{"why":"Provides the height-scan traversability heuristic baseline in the MPPI planning comparison.","marker":"[30]"},{"why":"Defines the Model Predictive Path Integral control framework that the FDM is integrated into.","marker":"[31]"},{"why":"Supplies the sim-to-real system identification approach that underlies synthetic data generation for legged locomotion.","marker":"[37]"},{"why":"Supplies the observation noise augmentation used on synthetic samples to make the FDM resilient to sensor noise.","marker":"[40]"},{"why":"Provides the real-world deployment and sensor data sources used for fine-tuning on snow, forest, and pavement.","marker":"[38]"},{"why":"Supplies the factor-graph pose estimation used as ground truth for real-world training labels.","marker":"[39]"}],"fun_headline_variants":["Learned model cuts robot position error by 41%","Zero-shot planner uses learned dynamics for safer navigation","Perceptive FDM boosts ANYmal success rate by 27%","No cost tuning: learned forward model enables safe robot planning","Sim-to-real transfer: learned model improves navigation on rough terrain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The height scan is encoded once at the start of the rollout, so the model must assume that a single static, partly occluded view of the terrain contains all geometry relevant to the next five seconds of motion.","fun_headline_variants_meta":{"raw":{"variants":["Learned model cuts robot position error by 41%","Zero-shot planner uses learned dynamics for safer navigation","Perceptive FDM boosts ANYmal success rate by 27%","No cost tuning: learned forward model enables safe robot planning","Sim-to-real transfer: learned model improves navigation on rough terrain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000451,"raw_usage":{"total_tokens":2345,"prompt_tokens":1091,"completion_tokens":1254,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":707,"completion_tokens_details":{"reasoning_tokens":1172}},"tokens_in":707,"tokens_out":1254,"duration_ms":10081,"temperature":1.0,"reasoning_tokens":1172,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:55:16.778915+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained FDM through a course where a low obstacle is hidden behind a ridge for the first two prediction steps and only becomes visible from the robot's onboard camera later; if the predicted poses keep passing through the obstacle with low risk, the single-scan assumption fails. A quantitative version would measure rollout error against a variant that re-encodes the height scan at every step.","supporting_citations":[{"cited_title":"Learning forward dynamics model and informed trajectory sampler for safe quadruped navigation","cited_arxiv_id":null,"evidence_quote":"Provides the first learned FDM for quadruped navigation and the perceptive baseline against which the 41% position-error improvement is measured."},{"cited_title":"Rough terrain navigation for legged robots using reachability planning and template learning","cited_arxiv_id":null,"evidence_quote":"Provides the height-scan traversability heuristic baseline in the MPPI planning comparison."},{"cited_title":"Information theoretic mpc for model-based reinforcement learning","cited_arxiv_id":null,"evidence_quote":"Defines the Model Predictive Path Integral control framework that the FDM is integrated into."},{"cited_title":"Learning agile and dynamic motor skills for legged robots","cited_arxiv_id":null,"evidence_quote":"Supplies the sim-to-real system identification approach that underlies synthetic data generation for legged locomotion."},{"cited_title":"Boxi: Design Decisions in the Context of Algorithmic Performance for Robotics","cited_arxiv_id":null,"evidence_quote":"Provides the real-world deployment and sensor data sources used for fine-tuning on snow, forest, and pavement."}],"review_version":1}