{"id":"9c771c2f-f125-494d-b993-d94443a0c28d","arxiv_id":"2505.10973","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A single offline-trained attention model, using no robot-specific encoding, transfers walking and stair-climbing across multiple quadruped robots, including hardware runs.","lead":"This paper trains one offline locomotion policy from expert demonstrations on several quadruped robots and shows it can walk and climb stairs on robots it never saw during training. The framework uses only joint and sensor data, with no robot-specific identifiers, so the same network may transfer across different robots and terrains.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training procedure carries recurrent hidden states across unrelated batches, so the claimed zero-shot policy is not the product of the described method; this undermines the reproducibility of the central claim.","rationale":"The reader's weakest assumption concerned the unstated diversity of the offline expert datasets. That is a legitimate concern, but the paper contains a more specific, self-contained technical flaw in the training procedure that is testable and directly affects whether the described method can produce the reported policy. In Section 2.4, the authors state that hidden states are preserved across epochs after warmup, even though each epoch samples a fresh random batch. This is inconsistent with the standard BPTT setup and with the zero-initialization used during evaluation and hardware deployment. If taken literally, the trained policy's initial recurrent state is drawn from a different distribution than at test time, so the reported zero-shot results may not reflect the architecture's ability to generalize from clean episode starts. The paper also defers key evidence to an absent supplementary material (hardware deployment details, additional analyses), which further limits verifiability. A conditional acceptance is appropriate: the authors should correct the training loop description and provide code or clear pseudocode, alongside trial-level evaluation data and the promised supplementary material, before the central claim can be trusted. My concern is not a substitute for the reader's data-diversity concern but a more pointed internal inconsistency that can be settled by a straightforward re-implementation.","tokens_in":10882,"tokens_out":11649,"duration_ms":125120,"concrete_test":"Re-implement the training loop with the standard practice of resetting each trajectory's hidden state to zero at the start of every epoch and at every true episode boundary (not at padding), removing the cross-epoch carryover. Then re-run the five settings in Table 1 and the Go1 hardware deployment. If the zero-shot successes (e.g., Go1 climbing 21 cm stairs in Setting 5) disappear or degrade, the reported generalization depends on the nonstandard hidden-state initialization and the central claim is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 2.4 describes a training loop in which each 'epoch' samples a single random batch of 400 trajectories, and after a 50-epoch warmup the recurrent hidden states are preserved across epochs: 'For e > Ew, hidden states are preserved, enabling continuity across (padded) episodes.' Because a new random batch is drawn each epoch, the initial hidden state for a trajectory in epoch e+1 is the final hidden state of an unrelated trajectory from epoch e. Standard truncated BPTT requires resetting hidden states at true episode boundaries. If the implementation follows the text, the model is trained under a distribution of initial recurrent states that differs sharply from the zero initialization used at deployment on real robots, so the reported zero-shot transfer to Go1 and Stoch5 cannot be attributed to the method as written. This is load-bearing because the paper's central claim is that behavior cloning from offline data, with the specified architecture and training, yields a deployable generalist policy; a flawed recurrent-state handling makes the reported results an artifact of an unspecified and likely incorrect training procedure rather than evidence for the claimed generalization mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GRoQ-LoCO, an attention-based behavior-cloning policy trained entirely offline on expert demonstrations of flat-ground and stair-climbing locomotion from multiple quadruped robots. The policy consumes only proprioceptive observations and command velocities, with no robot-specific identifiers. The authors claim zero-shot generalization to unseen robots (including hardware deployment on Unitree Go1 and Stoch5) and to out-of-distribution terrains such as higher stairs and slopes, across five training configurations with uneven distribution of skills across robots.","tokens_in":11121,"tokens_out":4911,"duration_ms":47255,"significance":"If the claims hold, the work would be a useful advance: it demonstrates that a single offline-trained policy can serve as a generalist controller across multiple quadruped morphologies and terrains without test-time adaptation or robot encoding, which goes beyond prior single-embodiment offline locomotion works. The framework is simple and the idea of behavior fusion from periodic and non-periodic gaits is appealing. However, the current evidence is not conclusive: the evaluation is largely binary, the dataset is underspecified, and a training-procedure detail appears internally inconsistent. The paper also includes a useful comparison with robot-encoding variants in the appendix, which supports the claim that explicit encodings hurt transfer.","major_comments":[{"comment":"The training procedure as written passes recurrent hidden states from a randomly sampled batch at epoch e to the next random batch at epoch e+1. Since each epoch samples a new batch of 400 trajectories and trajectories are padded to Nmax, the initial hidden state for a trajectory in epoch e+1 is the terminal hidden state of an unrelated trajectory from the previous epoch. This creates a training-time distribution of initial recurrent states that does not match the zero initial state used at deployment, and it is not the standard truncated BPTT described in the same section. The reported zero-shot results therefore cannot be attributed to the method as specified; please either correct the implementation description or retrain with proper episode-aligned hidden-state resetting and confirm the results.","section":"Section 2.4"},{"comment":"The central generalization claims rest on binary success/failure evaluations with no repeated trials, no success rates, and no statistical analysis. For example, Table 1 reports a single check/cross per robot per stair height, and Figure 7a similarly reports a single binary outcome per slope condition. Without variance estimates or repeated runs, it is impossible to distinguish genuine zero-shot transfer from chance or from favorable single runs. Please report aggregate statistics, such as mean success rates over multiple seeds and trials, with error bars or confidence intervals.","section":"Section 3.1 (Table 1) and Figure 7a"},{"comment":"The expert demonstrations are generated by two RL controllers that are never specified: no reward functions, domain randomization, terrain distributions, episode lengths, or data volumes per robot are given. Since the entire method is behavior cloning from these datasets, the generalization claim cannot be assessed or reproduced without this information. Please include expert training details, dataset statistics, and a description of the diversity across robots and terrains, or make the dataset publicly available with full documentation.","section":"Section 2.1"},{"comment":"The hardware deployment results on Go1 and Stoch5 are anecdotal ('demonstrated zero-shot transfer', 'robust walking') with no quantitative measurements such as success rates over repeated runs, gait frequency, joint torques, or failure cases. As this is the paper's most direct evidence for sim-to-real transfer, it needs to be quantified and compared against baselines or at least reported with clear metrics and experimental protocol.","section":"Section 3.2"}],"minor_comments":[{"comment":"Table 6 reports 'Embedding Dimension (emb dim) = 64' while Appendix D.1 states 'emb dim = 128', and Table 6 describes a 2-layer MLP head while Appendix D.1 says a 3-layer MLP; please reconcile these values so the architecture is reproducible.","section":"Table 6 and Appendix D.1"},{"comment":"The text says 'The most recent k outputs are aggregated' but k is never defined; the window size W=100 appears only in Appendix D.2. Define k in the main text and ensure consistent use of k versus W.","section":"Section 2.2"},{"comment":"The caption refers to 'GROQLoco' while the model is consistently named GRoQ-LoCO elsewhere; please standardize the name.","section":"Figure 2 caption"},{"comment":"The adaptive loss uses 'exp(-log σ)' which equals 1/σ; please simplify the expression or explain the intended adaptive weighting, and specify the initialization of the learnable parameter σ.","section":"Section 2.3"},{"comment":"The text refers to 'Stoch5(ZO)' in Figure 8 and Figure 9 captions, which appears to be a typo for 'ZS' (zero-shot).","section":"Appendix B"},{"comment":"The attention pattern figures appear as placeholders with no actual plots; please include the figures or remove the section, as the current presentation cannot be evaluated.","section":"Appendix D.3"},{"comment":"Reference [9] has an incomplete author field (a missing name before 'Gajewski'); please correct it.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The hidden-state training issue in Section 2.4 is the most serious technical problem: if the implementation truly follows the text, the reported results may be artifacts of an invalid recurrence handling. The experimental evaluation also falls well short of the standard for a robotics journal, being binary and unrepeated. I believe the central idea is worth pursuing, but the authors need to fix the training procedure, provide dataset and expert-policy details, and substantially upgrade the empirical evaluation before this can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper reports a single behavior-cloned policy trained on offline demos from a few quadrupeds that transfers zero-shot to unseen robots and terrains, with hardware demos on Go1 and Stoch5. If true, that is a useful step for multi-embodiment locomotion, and it is new relative to the cited work: GenLoco is flat-velocity only, MorAL uses online RL with adaptation, DiffuseLoco is single-robot. The architecture is sensible (attention + GRU + adaptive loss), the choice to avoid any robot-specific encoding is clean, and the cross-robot training settings with uneven skill distribution are a nice experimental idea. Appendix B's comparison to a robot-encoding variant is a useful control.\n\nThe soft spots are real and some are serious. First, the evaluation is a binary success/failure table with no repeated trials, no error bars, no success rates, and no quantitative measures of stability or traversal quality. The hardware deployment is anecdotal. This alone would require major revision. Second, the data pipeline is underdescribed: no reward functions for the expert RL controllers, no domain randomization details, no terrain variation, no dataset size. Without that, the zero-shot generalization claim is not interpretable. Third, the training loop in Sec 2.4 is problematic. After the warmup, hidden states are preserved across epochs, but each epoch samples a completely new batch. That means the initial hidden state for a trajectory in epoch e+1 is the final hidden state of an unrelated trajectory from epoch e. At deployment, the hidden state is zero-initialized. This is a serious train/test mismatch. If this is what was actually done, the reported zero-shot results cannot be attributed to the method as described; if it is a writing error, the text must be corrected. Either way, reproducibility is compromised. Fourth, there are no baselines: no single-robot policies, no standard BC without attention, no comparison to existing multi-embodiment methods. The robot-encoding comparison is a start but not enough. The reader's note about inconsistent robot-mode labels is a false alarm—modes vary by setting by design.\n\nThe core idea is plausible and worth pursuing, and the problem is important. But the current paper is not yet convincing. It deserves a serious referee because a substantiated version would be a real advance, and the authors should get a chance to fix it. My recommendation: send to peer review, expect major revision—repeated trials with quantitative metrics, full data generation details, a corrected training description, and baselines.","headline":"A zero-shot multi-robot locomotion claim worth taking seriously, but the evidence is too thin and the recurrent training loop as written is broken—needs a proper rewrite and repeated-trial evaluation.","tokens_in":11639,"tokens_out":4656,"would_cite":false,"duration_ms":48901,"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":"A single offline-trained behavior-cloning policy controls multiple quadruped robots it never encountered during training, walking on flat ground and climbing stairs, with no robot-specific encoding and no test-time adaptation.","keywords":["behavior cloning","quadruped locomotion","zero-shot transfer","offline learning","generalist policy","multi-embodiment","attention network","terrain generalization"],"falsifier":"Remove stair-climbing demonstrations from the training set for all robots and evaluate the trained policy on stairs; if stair traversal still succeeds, behavior fusion is not the mechanism behind zero-shot transfer. The paper's Setting 4 (stair data from Go1 only) already shows B2 and Aliengo failing at 17 cm, so a fully stair-free dataset would confirm that diversity in training data is what carries the transfer.","tokens_in":10712,"feed_emoji":"🤖","tokens_out":6729,"duration_ms":64064,"temperature":0.7,"pith_summary":"The paper proposes GRoQ-LoCO, a generalist quadruped locomotion controller trained entirely offline by behavior cloning on expert demonstrations of two skills—flat-ground walking and stair climbing—collected from several different quadruped robots. The central claim is that dataset diversity, across both robot morphologies and locomotion behaviors, is what enables a single policy to transfer zero-shot to robots it never saw, including real hardware, with no robot-specific encoding and no test-time adjustment. The authors argue that offline imitation alone, without online RL or morphology conditioning, can fuse periodic and non-periodic behaviors into one network. If correct, this would make locomotion policies reusable and safer to train, since all data is collected in simulation and the resulting policy runs in real time on an embedded processor.","feed_headline":"One offline-trained policy controls quadrupeds it has never seen","feed_subtitle":"Zero-shot flat walking and stair climbing across robots, with no robot-specific data or test-time adaptation","key_machinery":"An attention-augmented recurrent policy trained by behavior cloning forms the mechanism. Each proprioceptive observation (joint angles and velocities, previous two actions, gravity vector, angular velocity, commanded velocity) passes through a linear encoder with LayerNorm; sinusoidal positional embeddings are added, and multi-head self-attention runs twice—once over a sliding window of recent observations and once over a sliding window of recent GRU hidden states—with a final MLP head producing joint actions. Two design choices carry the argument: the absence of any robot identifier, forcing the policy to infer embodiment purely from dynamics, and an adaptive per-joint loss (a robust Huber-like term with a learned per-dimension variance) that re-weights joints according to modeling difficulty. The attention analyses are used to claim that the heads learn terrain-conditioned, robot-independent features.","core_discovery":"The central claim is that a single policy trained by behavior cloning on offline expert demonstrations of two behaviors—periodic flat-terrain gait and non-periodic stair traversal—collected across several quadruped robots, generalizes zero-shot to unseen robot morphologies and to terrains (higher stairs, slopes) absent from training. The policy reads only proprioceptive observations (joint positions, velocities, previous actions, gravity vector, angular velocity, commanded velocity) and outputs actions at low latency on an embedded computer. No robot identifier, no morphology descriptor, and no test-time optimization is used. The paper reports that training configurations with full diversity (Setting 5, both behaviors and three morphologies) give the strongest zero-shot transfer, that stair-only robots extrapolate to out-of-distribution step heights, and that hardware deployment on the Unitree Go1 and Stoch5 succeeds without fine-tuning.","pith_inferences":["If the result holds beyond the tested kinematic range, behavior-cloned generalist policies could let the community pool demonstration data across robot models the way manipulation datasets do today; the paper's own limitation section flags the need to test more diverse leg geometries.","The learned per-joint variance in the adaptive loss doubles as a diagnostic of which joints are hardest to model per terrain, which could guide where to collect more demonstrations.","A direct test of the behavior-fusion hypothesis would be a training set that excludes one behavior entirely; the paper's Setting 4 partially does this and shows degraded transfer, suggesting that fusion is not automatic but depends on the diversity mix.","Combining this architecture with a lightweight exteroceptive input, as the authors list for future work, is a natural next step to move from reactive terrain handling to planned navigation."],"forward_implications":["A single policy trained once in simulation can be deployed across a fleet of different quadruped models without retraining, fine-tuning, or robot identifier.","Stair-climbing ability transfers to robots that only contributed flat-ground demonstrations, indicating that periodic-gait experience alone generalizes to non-periodic terrain.","Policies trained with full diversity extrapolate beyond their training range: robots trained on 17 cm stairs later climb 25–29 cm steps.","Novel terrains absent from the training set (slopes up to 40°, both smooth and rough) are handled zero-shot by several robots.","The same policy runs in real time on an embedded computer and transfers to physical hardware without test-time optimization."],"supporting_citations":[{"why":"Defines the paradigm of pooling demonstration data across different robot embodiments in manipulation; the model this paper adapts from arms to legs.","marker":"[3]"},{"why":"Sets the template for scale-driven generalist policies in manipulation; motivates the offline dataset approach.","marker":"[1]"},{"why":"Demonstrates offline diffusion-based locomotion with zero-shot sim-to-real on a single robot; the baseline this work extends across embodiments.","marker":"[6]"},{"why":"Shows morphology randomization to train one policy across quadrupeds on flat terrain; the key prior comparison for robot-agnostic control.","marker":"[18]"},{"why":"Adds an adaptive dynamics module for morphology generalization; a prior approach that still relies on online RL.","marker":"[20]"}],"fun_headline_variants":["One offline policy walks any quadruped, zero-shot","Offline-trained controller zero-shots across quadrupeds","Offline data only: one policy for all quadrupeds, seen or not","Unseen quadrupeds walk and climb with this offline-trained policy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The expert demonstrations are generated by two simulation RL controllers whose reward functions, domain randomization, terrain variation, and dataset size are not specified; if these demonstrations are too narrow or too simulation-specific, the zero-shot transfer seen in the paper would not occur.","fun_headline_variants_meta":{"raw":{"variants":["One offline policy walks any quadruped, zero-shot","Offline-trained controller zero-shots across quadrupeds","Offline data only: one policy for all quadrupeds, seen or not","Unseen quadrupeds walk and climb with this offline-trained policy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001083,"raw_usage":{"total_tokens":4565,"prompt_tokens":1021,"completion_tokens":3544,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":3471}},"tokens_in":637,"tokens_out":3544,"duration_ms":22897,"temperature":1.0,"reasoning_tokens":3471,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:00:08.567572+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Remove stair-climbing demonstrations from the training set for all robots and evaluate the trained policy on stairs; if stair traversal still succeeds, behavior fusion is not the mechanism behind zero-shot transfer. The paper's Setting 4 (stair data from Go1 only) already shows B2 and Aliengo failing at 17 cm, so a fully stair-free dataset would confirm that diversity in training data is what carries the transfer.","supporting_citations":[{"cited_title":"Huang, Y","cited_arxiv_id":null,"evidence_quote":"Demonstrates offline diffusion-based locomotion with zero-shot sim-to-real on a single robot; the baseline this work extends across embodiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows morphology randomization to train one policy across quadrupeds on flat terrain; the key prior comparison for robot-agnostic control."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Adds an adaptive dynamics module for morphology generalization; a prior approach that still relies on online RL."}],"review_version":1}