{"id":"f5b3b7f8-1744-4213-a200-42116910da5d","arxiv_id":"2507.04039","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A transformer with body tokenization and consistent dropout generalizes to unseen leg damages and sensor noise while trained on limited dynamics and clean observations.","lead":"Robots trained in simulation often fail when a leg breaks or a sensor gets noisy. This paper proposes ROLT, a transformer design with body tokenization and dropout that keeps legged robots walking in several unseen failure and noise conditions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The action token's role as a data-dependent position encoding is confounded by the Module PE ablation; a global-state explanation for ROLT's cross-limb generalization remains viable.","rationale":"The reader's weakest assumption listed both the representativeness of held-out failures and the action-token position mechanism; I focus on the latter because it is more concrete and testable. The empirical OOD results are promising and the fall-time table is strong, but the paper's main contribution is the architectural principle of body tokenization. Without a clean ablation separating the action token's positional role from its role as a global action or coordination token, the claim that body tokenization supports knowledge sharing across limbs could be an artifact of added global context. This is not a defect in the experiments themselves, and it is addressable by a few additional training runs, so it does not change the reader's conditional verdict.","tokens_in":9935,"tokens_out":7147,"duration_ms":89366,"concrete_test":"Run the A1 stiffness and hexapod damage tests with three additional ablations: (1) keep the action token and add data-independent learnable PE to leg tokens; (2) keep the action token but shuffle the order of its per-leg action components for one leg at test time; (3) replace the action token with a global state token (e.g., mean or concatenated body state with no leg ordering) that carries no positional information. If (1) matches ROLT or (2) does not degrade performance, the action token is not serving as a position encoding and the mechanism story fails; if (3) matches ROLT, the benefit is global context rather than limb position.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-A claims that replacing data-independent position embeddings (PE) with an action token lets leg modules 'query their position information' and prevents overfitting to PE, and the Module PE ablation in Sec. IV-A is presented as evidence for this claim. However, Module PE differs from ROLT in two ways at once: it removes the action token and adds learnable per-leg embeddings. The action token is not only a positional query; it is also the only global token carrying the previous target angles of all joints. Since each leg token already contains its own 10-step joint history, the action token's contribution could be a global coordination or context signal rather than leg identity. If that is the case, the hexapod damage results do not demonstrate the proposed position-query mechanism, and the claimed link from body tokenization to cross-limb knowledge transfer is under-identified. This does not invalidate the empirical generalization result, but it leaves the central architectural explanation without decisive support.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ROLT, a transformer-based locomotion policy whose two main design elements are body tokenization (each leg is a token, plus an action token that replaces data-independent position embeddings) and consistent dropout (feature dropout and patch-PE dropout applied with the same mask during rollout and update). The policy is trained with PPO in IsaacGym on the Unitree A1 and a hexapod, using only a few damaged conditions and clean elevation observations. Evaluation covers simulated limb stiffness changes, unseen joint failures on the hexapod, low- and high-frequency elevation-map noise, and three real-world perturbations on a Unitree Go1. The paper claims that ROLT generalizes across limbs and to unseen perceptual noise better than MLP, Roll-Drop, RMA, and ROAM baselines.","tokens_in":10051,"tokens_out":5048,"duration_ms":58369,"significance":"If the reported results hold, this is a useful empirical contribution: it demonstrates that a transformer with body-level tokenization can transfer fault-tolerant behavior across limbs after training on only a handful of damage conditions, and that a consistent-dropout scheme can improve robustness to observation corruption without explicit noise injection during training. The hexapod damage results in Table II, which include means and standard deviations, are concrete evidence for the dynamics-generalization claim, and the real-world transfer experiments add practical credibility. The central architectural interpretation, namely that the action token acts as a data-dependent position encoding, is plausible but is not decisively supported by the current ablations; this should be fixed before the mechanism claim is accepted.","major_comments":[{"comment":"The Module PE ablation changes two variables at once: it removes the action token and it adds learnable per-leg position embeddings. The action token is not only a positional query; it is also the only token that carries the previous target angles of all joints, i.e., a global action-context signal. Therefore the large gap between ROLT and Module PE in the hexapod damage test (Table II) could be explained by the loss of this global context, or by the introduction of per-leg embeddings, rather than by the absence of a data-dependent position-query mechanism. I agree with the stress-test concern that this is a real confound. Please add ablations that separate these factors, such as ROLT with per-leg learnable PE added in addition to the action token, or Module PE with a non-positional global token carrying the same action vector, and/or attention analyses showing that leg tokens use the action token specifically for position information.","section":"Sec. III-A, Eq. (2) and Sec. IV-A, Module PE ablation"},{"comment":"The A1 stiffness test and both perceptual-noise tests are presented only as plots, with no numeric tables, no number of seeds, no standard deviations, and no test-episode counts. These experiments are load-bearing for the paper's claims of dynamics generalization and perceptual robustness, so without variance information the reader cannot assess whether ROLT's advantage is statistically meaningful or whether the curves reflect one or a few favorable runs. Please report mean/std tables and seed counts for the stiffness and noise experiments, as is already done for the hexapod damage test.","section":"Sec. IV-A, Fig. 3 and Sec. IV-B, Figs. 6-7"},{"comment":"The real-world experiments report means and standard deviations but do not state the number of trials per condition, and they compare against only RMA. Given the small distances involved (0.0-4.8 m over 10 s) and the strength of the claim that ROLT generalizes in real-world deployment, this table needs the number of runs per condition and ideally per-trial results. Without these details, the real-world conclusion is difficult to evaluate.","section":"Sec. IV-C, Table III"}],"minor_comments":[{"comment":"The figure captions appear to be swapped: the text says low-frequency noise results are shown in Figure 6 and high-frequency noise results in Figure 7, but the captions read \"Fig. 6. High-frequency noise test results\" and \"Fig. 7. Low-frequency noise test results.\"","section":"Sec. IV-B, Figs. 6-7"},{"comment":"The fall-time entry for HRC is truncated as \"6.74±3.\"; please provide the full value.","section":"Table II"},{"comment":"There are several typos and grammar issues, including \"prioprocetion,\" \"hexpodal,\" \"contary,\" \"wired gait\" (should be \"weird gait\"), \"Gussian,\" \"a extremely,\" and \"both ROLT maintain.\" A careful proofreading pass is needed.","section":"Throughout"},{"comment":"The attention visualization is described as averaging \"the attention score of the map patches on each leg token,\" but the text does not specify which attention heads or layers are averaged; please specify the exact aggregation to make the analysis reproducible.","section":"Sec. IV-D, Fig. 8"},{"comment":"The foot airtime reward term is listed without a coefficient, while all other reward terms have one; please clarify whether it has a coefficient or is intentionally unweighted.","section":"Table I"},{"comment":"The text says that consistent dropout applies the same mask during rollout and update, but it is not stated whether the same mask is shared across parallel environments or across time steps; please clarify the implementation.","section":"Sec. III-C"}],"recommendation":"major_revision","confidential_remarks":"The empirical result is interesting and likely publishable after revision, but the central mechanism claim about the action token is under-identified by the current Module PE ablation. I would ask the authors to either add disentangling ablations or soften the mechanistic interpretation. The missing numeric reporting for the stiffness and noise experiments is also important for a fair assessment. No citation or ethics concerns beyond what is already in the report."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: the hexapod damage test is a strong empirical result, but the paper's explanation for why it works is not established by the experiments. Trained with failures on just three joints (plus healthy), ROLT walks the full 20 seconds in all nine unseen joint-failure conditions, with mean distances around 13-16 m; every baseline falls in at least some conditions and usually walks less than half as far. That is a concrete, falsifiable generalization result, and it is the core value of this paper.\n\nWhat's actually new: an action token in place of learnable per-leg position embeddings, plus dropout applied to the patch position embeddings. The PE dropout appears to help the policy ignore corrupted terrain patches, and the attention maps in Fig. 8 give a nice qualitative view of that behavior. The consistent dropout mechanism itself comes from Hausknecht and Wagener, so the novelty is the combination and the scale, not the base technique.\n\nThe soft spots are real but addressable. The stress-test note has it right: the Module PE ablation changes two things at once. It removes the action token and adds per-leg embeddings, so the performance gap could come from losing the global action context rather than losing a positional query. Since each leg token already contains its own 10-step history, the action token's unique content is the previous target angles of all joints - that's a global coordination signal, not obviously a positional one. The paper never analyses attention between the action token and leg tokens, so the mechanism claim is speculative. The stiffness test and noise tests are shown only as plots, with no numeric tables, error bars, or seed counts. The real-world experiments use only RMA as a baseline and cover just three scenarios; the results are promising but thin. And there is no code or data release, so reproducing the sim results is not possible.\n\nFor the right reader - someone working on legged locomotion robustness or transformer policies for control - this paper is worth a serious look. I would send it to peer review rather than desk-reject. The hexapod result alone justifies referee time. The main revision request should be a cleaner ablation that isolates the action token's role, plus statistics for the plotted results and ideally open-source code. I wouldn't cite it in its current form, but I'd revisit after revision.","headline":"A strong hexapod damage result makes this worth reviewing, but the action-token mechanism is under-identified and several plots lack statistics.","tokens_in":10589,"tokens_out":4183,"would_cite":false,"duration_ms":44156,"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 transformer policy trained on clean walking and one leg failure can keep a robot walking through unseen leg failures and sensor corruption.","keywords":["legged locomotion","out-of-distribution generalization","transformer policy","body tokenization","consistent dropout","fault tolerance","reinforcement learning","perceptual noise"],"falsifier":"A direct test would be to train ROLT exactly as described, with healthy walking plus one locked leg, and then evaluate it on failures outside the tested set: a hip joint locked, a leg with two joints disabled, partial motor weakening rather than full lock, or a change in payload. If the robot falls within seconds on any of these, the claim that one leg's fault experience transfers to arbitrary new limb failures would be falsified. Similarly, injecting noise into proprioception while keeping the elevation map clean would test whether the robustness mechanism really re-anchors on trusted body state or only on the specific perception corruption studied.","tokens_in":9688,"feed_emoji":"🦿","tokens_out":7131,"duration_ms":69141,"temperature":0.7,"pith_summary":"ROLT is a transformer policy for legged robots that is claimed to keep walking in situations that differ from training, without domain randomization or an online adaptation module. The controller is built around two ideas: body tokenization, in which each leg is a token that shares attention with the other legs and queries an action token instead of a learned position embedding, and consistent dropout, in which the same neuron mask and patch-position mask are applied during rollout and update. Trained only on healthy walking plus a single limping leg on a quadruped, or a handful of joint failures on a hexapod, the policy is reported to stay upright through unseen leg failures and to tolerate elevation-map corruption when trained on clean observations. The paper's claim is that these network-structure choices, rather than more elaborate training schemes, are what give the controller its out-of-distribution generalization.","feed_headline":"A robot policy trained on one limp walks through unseen failures","feed_subtitle":"Leg tokens and consistent dropout let a transformer keep walking through novel damage and noisy sensors.","key_machinery":"The load-bearing mechanism is the tokenization of the robot body into leg tokens, each containing ten steps of joint positions, joint velocities, previous actions, velocity command, and IMU history, projected by a shared MLP; a linear embedding of the previous action forms an action token. The action token replaces learnable position embeddings as the source of position information, so legs can query their roles through attention, allowing experience to transfer from one limb to another. Robustness to perceptual noise is carried by two consistent dropout operations in the two-layer encoder: feature dropout at rate 0.1 after attention and feed-forward layers, and patch position-embedding dropout at rate 0.75. Consistency means the same dropout mask is used at rollout and at update, which allows a high dropout rate to be used without destabilizing policy-gradient training.","core_discovery":"On the paper's own terms, the central discovery is that a transformer can transfer fault-tolerance experience across limbs if position information is made data-dependent. Instead of adding fixed position embeddings to leg tokens, ROLT appends an action token encoding the previous action, and each leg token infers its spatial role through attention to this token. In the hexapod damage test, this design keeps the robot from falling in all nine unseen joint-failure conditions over 20 seconds, while the baselines fall in most of the same tests. The consistent-dropout half of the claim is that dropping features and patch position embeddings, with the same mask during rollout and update, produces a policy that handles high- and low-frequency perceptual noise even though it was trained with clean elevation maps. Attention visualizations are used to argue that position-embedding dropout makes the network shift its focus to trustworthy patches instead of anchoring to learned positions.","pith_inferences":["A natural extension the paper does not test is whether body tokenization transfers to any control problem with modular repeated structure, such as multi-finger hands, multi-robot teams, or configurable limbs, where combinatorial generalization across modules is desired.","The position-embedding dropout result suggests a general recipe for sensor-fusion robustness: make the network doubt the spatial anchoring of one modality so it re-anchors using a trusted modality; the paper demonstrates this with elevation maps, but it may extend to depth cameras, lidar, or tactile arrays.","A stress test the paper does not run is corruption of the proprioceptive channel itself; because the attention re-anchors on proprioception, noise injected into joint encoders or IMUs could be the first failure point.","The training set is framed as limited, but the experiments cover only discrete lock-or-disable joint failures; whether the mechanism extends to continuous parameter drifts, such as gradually weakening motors or changing payload, is left open."],"forward_implications":["If the central claim is right, a locomotion policy does not need to be trained in every failure mode; one or a few fault experiences can generalize to other limbs of the same body plan.","The same architecture, with the number of leg tokens changed, transfers from a four-legged to a six-legged robot, suggesting the controller is scalable to morphologies built from repeated limb modules.","Because noise robustness comes from a dropout strategy and attention rather than from adding Gaussian noise to training observations, policies can be trained on clean data and still deployed where sensors are corrupted or missing.","The hexapod's redundant joints can be exploited by the controller to keep walking when motors are disabled, so the method could be used for fault-tolerant operation on robots whose training time must be kept short."],"supporting_citations":[{"why":"Supplies the RMA baseline with an adaptation module that must be beaten in the dynamic OOD tests.","marker":"[3]"},{"why":"Supplies the ROAM baseline that selects behavior modules and serves as the single-life RL comparison.","marker":"[10]"},{"why":"Provides the attention mechanism on which the entire ROLT architecture is built.","marker":"[11]"},{"why":"Supplies the patch tokenization and position-embedding conventions for the elevation map that ROLT modifies.","marker":"[17]"},{"why":"Supplies the consistent-dropout idea of using the same dropout mask at rollout and update to stabilize training under larger dropout rates.","marker":"[23]"},{"why":"Supplies the Roll-Drop baseline and the observation that dropout during rollout can simulate sensor noise.","marker":"[24]"},{"why":"Supplies the limb-module graph view of robot morphology that motivates tokenizing each leg separately.","marker":"[25]"},{"why":"Supplies the policy-gradient algorithm used to train all policies in the experiments.","marker":"[26]"},{"why":"Supplies the massively parallel simulation training setup that makes the large-scale training runs feasible.","marker":"[27]"}],"fun_headline_variants":["Transformer with body tokens walks through unseen failures","Consistent dropout makes transformer robust to sensor noise","Robust transformer generalizes locomotion to new conditions","Action tokens teach robot legs to handle novel damage"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the held-out failures tested (three other limping legs, nine joint disables, and three real-world perturbations) stand in for the entire class of out-of-distribution dynamic and perceptual changes; the paper does not formally characterize that class, so if the success is specific to these particular failures or to the simulation distribution, the broad robustness claim would not follow.","fun_headline_variants_meta":{"raw":{"variants":["Transformer with body tokens walks through unseen failures","Consistent dropout makes transformer robust to sensor noise","Robust transformer generalizes locomotion to new conditions","Action tokens teach robot legs to handle novel damage"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000359,"raw_usage":{"total_tokens":1926,"prompt_tokens":909,"completion_tokens":1017,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":959}},"tokens_in":525,"tokens_out":1017,"duration_ms":9405,"temperature":1.0,"reasoning_tokens":959,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:57:17.207626+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to train ROLT exactly as described, with healthy walking plus one locked leg, and then evaluate it on failures outside the tested set: a hip joint locked, a leg with two joints disabled, partial motor weakening rather than full lock, or a change in payload. If the robot falls within seconds on any of these, the claim that one leg's fault experience transfers to arbitrary new limb failures would be falsified. Similarly, injecting noise into proprioception while keeping the elevation map clean would test whether the robustness mechanism really re-anchors on trusted body state or only on the specific perception corruption studied.","supporting_citations":[{"cited_title":"Adapt On-the-Go: Behavior Modulation for Single-Life Robot Deployment","cited_arxiv_id":"2311.01059","evidence_quote":"Supplies the ROAM baseline that selects behavior modules and serves as the single-life RL comparison."},{"cited_title":"Consistent Dropout for Policy Gradient Reinforcement Learning","cited_arxiv_id":"2202.11818","evidence_quote":"Supplies the consistent-dropout idea of using the same dropout mask at rollout and update to stabilize training under larger dropout rates."},{"cited_title":"Roll-drop: accounting for observation noise with a single parameter,","cited_arxiv_id":null,"evidence_quote":"Supplies the Roll-Drop baseline and the observation that dropout during rollout can simulate sensor noise."},{"cited_title":"Learning modular robot control policies,","cited_arxiv_id":null,"evidence_quote":"Supplies the limb-module graph view of robot morphology that motivates tokenizing each leg separately."}],"review_version":1}