{"id":"f309e3f7-1772-495d-945a-3143705f0504","arxiv_id":"2412.04273","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A quadruped robot acquires walking, jumping, running-like, and standing skills using only the output of a video classifier trained on wild-animal videos as its reinforcement learning reward.","lead":"A quadruped robot learned to walk, jump, and stand still by watching thousands of wild-animal videos from nature documentaries, using a video classifier as the reward signal. The method needs no motion-capture data or hand-designed per-skill rewards, and the learned skills transferred to a real robot.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reward grounding is unvalidated: the video classifier is never evaluated on robot renders, and the paper's own Fig. 4 shows it can be fooled by a two-leg walking-in-place artifact.","rationale":"Both the reader and this pass identify the same load-bearing assumption: the video classifier must transfer zero-shot from wild-animal footage to synthetic robot renders and must not be exploitable. The paper's Fig. 4 is direct evidence that this assumption can fail, and no experiment in the paper measures the classifier's behavior on robot-domain inputs. This is more central than the walking/running indistinctness or the lack of released code: even if the metrics were perfect, the learned behaviors would not correspond to animal action classes if the reward is driven by artifacts. The paper does provide real evidence—the curated classifier is necessary (no-curating ablation fails), and the policy attains forward velocity and jumping height—so I would not reject or mark unverdictable. The appropriate disposition remains conditional: accept the existence of a pipeline that can yield locomotion-like behaviors, but not the stronger grounding claim, until the reward is directly validated on robot videos. The proposed confusion-matrix test would settle whether the concern actually lands.","tokens_in":16270,"tokens_out":5285,"duration_ms":57779,"concrete_test":"Using a scripted PD controller in IsaacGym, generate a held-out set of robot videos with known ground truth: standing still, walking in place, forward trot at low and high speed, jumping in place, and the two-leg shuffle from Fig. 4. Run the trained classifier on these videos and report the full 4-class softmax confusion matrix. If the classifier does not assign the correct skill the highest probability on the normal motions, or assigns 'walking'/'running' high probability to the two-leg shuffle or to a static robot with background motion, then the reward is demonstrably hackable and the central claim is not supported. If instead the classifier scores the scripted motions correctly and rejects the artifacts, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 defines the reward as the classifier probability fθ(x_robot, y) from Eq. (4), where x_robot is an 8-frame synthetic IsaacGym render. The classifier is trained exclusively on Animal Kingdom natural videos, and the paper provides no accuracy or confusion-matrix evaluation of this classifier on robot-domain videos. The only indirect evidence is the downstream policy metrics, but those can be maximized by reward hacking. Figure 4 documents exactly this: a policy that 'walks in place with two legs' is said to fool the video reward function, i.e., it obtains high classification reward without performing the commanded skill. This is not a peripheral failure; it demonstrates that the reward is not a faithful measure of the animal action concept on out-of-distribution robot inputs. If the classifier is responding to background, body pose, leg-motion artifacts, or camera configuration rather than the skill, then the learned 'walking', 'running', and 'jumping' behaviors are not grounded in animal motion. The real-robot deployment is qualitative and does not isolate the reward's validity. Thus the central claim is conditional on an unvalidated generalization step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces RLWAV, a pipeline in which a video action classifier trained on naturally occurring animal videos from the Animal Kingdom dataset is used as the scalar reward for a constrained reinforcement-learning policy that controls a Solo-12 quadruped in IsaacGym, after which the policy is transferred to the physical robot. The authors report simulation metrics across four skills, ablations of both the classifier and the policy-learning choices, and qualitative real-robot demonstrations. The central claim is that distinct locomotion skills can be acquired from wild animal videos without reference trajectories or skill-specific reward functions.","tokens_in":16515,"tokens_out":4762,"duration_ms":49750,"significance":"If the claim is upheld, the work is a useful step toward using internet-scale video corpora as reward sources for legged locomotion, complementing prior work on human-video-based manipulation rewards. Strengths include the use of a public dataset, multi-seed reporting, systematic ablations (classifier curation, model soup, camera position, constraint removal), and an honest discussion of failure cases and limitations. The main risks are that the learned reward is not demonstrated to be grounded in the intended animal actions when applied to synthetic robot renders, and that the real-robot evidence is qualitative; these gaps directly affect the strength of the central claim. There is no circularity concern because the reward is explicitly a learned classifier score and the evaluation metrics are independent, but the out-of-distribution validity of that classifier score is the load-bearing issue.","major_comments":[{"comment":"The reward used for RL is the classifier score fθ(x_robot, y) on synthetic robot videos, but the classifier is only trained on, and as far as reported only evaluated on, Animal Kingdom natural videos. The paper provides no accuracy, confusion matrix, or per-class reliability of this classifier on robot-domain inputs. This matters because the only validation of the reward is through downstream policy behavior, and the paper's own Figure 4 documents a failure mode in which a walking-in-place policy 'fool[s] the video reward function.' That example shows that high classifier reward can be obtained without performing the commanded skill, so the central claim that the learned behaviors are grounded in animal motion is conditional on an unvalidated out-of-distribution generalization step. I recommend adding a direct evaluation of the video classifier on robot renders (e.g., per-skill classification accuracy or reward-versus-behavior correlation) and, if possible, a reward-hacking analysis or a calibration against the quantitative skill metrics.","section":"§3.2, Eq. (4); §4.2, Fig. 4"},{"comment":"The real-robot experiments are qualitative: no measured forward speeds, jump heights, contact timings, or success rates are reported, and no video classifier scores or manual style ratings are taken on the real robot. As a result, the claim of 'successful transfer' and 'directly deploy' in §4.3 is not quantitatively supported. I ask for quantitative real-robot measurements (e.g., forward velocity, vertical displacement, and optionally stance/duty-factor estimates) and, if style is to be assessed, a small multi-rater study with reported inter-rater agreement.","section":"§4.3"},{"comment":"The distinction between walking and running is not established. In Table 1, walking velx = 23.0 ± 15.4 cm/s and running velx = 35.1 ± 12.9 cm/s overlap by more than one standard deviation, and the text concedes that the running policy does not produce flying phases and that on the real robot the two skills 'appear alike.' Since one of the four considered action classes is 'Running,' the abstract-level claim that the robot acquires skills corresponding to the action classes needs either a measurable gait criterion (e.g., duty factor, Froude number, flight-phase duration) or a softened claim that only three distinct behaviors are demonstrated.","section":"§4.2, Table 1; §4.3"}],"minor_comments":[{"comment":"There are several typos and minor awkward phrasings: 'activites' in §1, 'fishs' in §3.2, 'adaptatively optimized' in Appendix A.2, and the consistent spacing artifacts in 'RLW A V' throughout the captions. These should be corrected in a revision.","section":"§1, §3.2"},{"comment":"The alternative camera positions (Camera 1 through Camera 4) are described only as 'more extreme' and shown in a schematic; the exact placement relative to the robot, field of view, and distance are not specified, making the camera ablation difficult to reproduce. Please provide numerical camera parameters or a precise description in the appendix.","section":"Figure 4 and Table 2"},{"comment":"The adaptive reward normalization coefficients α_y and β_y are said to be 'adaptatively optimized' based on reward statistics across actors, but the precise update rule, initialization, and effect on the learned reward scale are not described. Since these coefficients are per-skill free parameters, a brief specification would improve reproducibility.","section":"Appendix A.2, Eq. (6)"},{"comment":"The style scores are presented as single numbers without stating how many raters evaluated the rollout videos, whether the ratings were blind to the skill command, or whether any inter-rater agreement was computed. This makes the style column difficult to interpret as evidence of skill quality.","section":"§4.1"},{"comment":"The claim 'without skill-specific reward functions' should be read carefully: the foot air-time constraint is applied uniformly across skills but is specifically designed to enable walking and running (Appendix B shows its removal destroys those skills). This is not a fatal issue, but the wording in §1 and §5 should acknowledge that the method still relies on locomotion-specific inductive biases.","section":"§3.3, Appendix A.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and addresses a timely question. My major-revision recommendation is driven by evaluation completeness rather than by any suspicion of methodological dishonesty: the central idea is promising, but the reward-grounding validation and real-robot quantification need to be strengthened before the stated claims can be accepted. I would not require perfection, but I do require that the authors either add the missing evaluations or explicitly narrow the claims they make."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Elliot — quick take on arXiv:2412.04273 (RLWAV). The genuinely new thing here is that a video classifier trained on internet wild-animal footage can serve as the sole task reward for learning quadruped locomotion skills — no reference trajectories, no per-skill reward shaping, no pose tracking. That claim mostly holds up in simulation and transfers to a real Solo-12. The paper is honest about what it can't do: walking and running look alike, running has no flying phase, and the behaviors are behind SOTA.\n\nThe strongest part is the experimental structure. Four seeds, informative ablations (data curation, model soup, camera position, constraints), a real-robot deployment, and a documented failure case where the robot walks in place on two legs and fools the video reward. That last point is the paper's biggest vulnerability. The reward is a classifier trained only on Animal Kingdom clips; it is never evaluated on robot renders. The only evidence that it generalizes is the downstream task metrics — but those can be maximized by hacking the classifier, and Fig. 4 shows exactly that happening. So the central claim — that the learned skills are grounded in animal motion — rests on an unvalidated zero-shot transfer step. It's the load-bearing assumption and the paper doesn't close the loop.\n\nThe other soft spots are more minor: real-world results are qualitative (no measured speeds or jump heights), style scores are manual with no variance or inter-rater checks, and the walking/running distinction is weak. The paper's own conclusion concedes the running behavior is limited. On the positive side, there is no circularity problem: the reward is an explicit classifier score and the evaluation uses independent metrics.\n\nWho is it for: anyone working on video-based reward learning, cross-embodiment transfer, or legged locomotion from observation. It deserves a serious referee. I'd recommend conditional accept: ask for quantitative real-robot measurements, a robustness check of the classifier on robot-domain videos (at minimum a confusion matrix or a background-correlation test), and a clearer separation of walking vs running. Code/models would help but aren't essential.","headline":"A genuine existence proof that wild-animal video classifiers can reward quadruped locomotion, but the reward grounding on robot renders is unvalidated and the walking/running distinction is weak.","tokens_in":17047,"tokens_out":2145,"would_cite":true,"duration_ms":22408,"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":"A video classifier trained on 8,791 wild-animal clips supplies the reward that teaches a quadruped robot to walk, jump, and stand still, without reference trajectories or per-skill reward functions; the policy transfers to a real robot.","keywords":["reinforcement learning","legged locomotion","video-based reward","cross-embodiment transfer","action recognition","quadruped robot","sim-to-real","animal videos"],"falsifier":"Feed the trained classifier robot clips with the body masked or blurred while the background and camera motion are preserved: if the 'Walking' or 'Running' probability stays high, the reward is tracking visual confounds rather than the skill. A direct correlation measurement between the classifier score and ground-truth forward velocity or foot air time across policy rollouts would likewise reveal how much of the reward is skill-grounded, and the two-leg walking-in-place failure reported in the paper is a concrete case where the score stayed high while the skill was absent.","tokens_in":16045,"feed_emoji":"🤖","tokens_out":10453,"duration_ms":92435,"temperature":0.7,"pith_summary":"This paper asks whether a legged robot can learn locomotion skills by watching internet footage of wild animals, and answers yes. The authors train a video classifier on 8,791 clips from a large animal-behavior dataset to recognize four actions — keeping still, walking, running, jumping — and then use the classifier's score on third-person videos of a simulated quadruped as the only task reward in reinforcement learning. Physical plausibility comes entirely from embodiment-agnostic constraints, not from skill-specific reward design. The resulting multi-skill policy transfers directly to a real Solo-12 robot, which trots forward on command, jumps in place, and stands still. The intended conclusion is that raw, cross-embodiment video data can substitute for hand-designed rewards and reference motions in locomotion learning.","feed_headline":"Robot learns walking and jumping from wild animal videos","feed_subtitle":"A video classifier trained on 8,791 nature clips provides the reward — no reference motions, no hand-designed rewards.","key_machinery":"The central object is the classifier-as-reward loop: a frozen video classifier $f_\\theta$ (a Uniformer-S transformer fine-tuned from Kinetics-400 pretrained weights, with random-convolution augmentations and weight-averaged model soups) turns pixels into a reward $R(s_t,a_t) = \\alpha_y f_\\theta(x_{\\mathrm{robot}}, y) + \\beta_y$ on frames captured every five simulation steps, where $y$ is the commanded skill and $\\alpha_y, \\beta_y$ are per-skill normalization terms. The second load-bearing piece is constrained reinforcement learning: Proximal Policy Optimization combined with Constraints as Terminations (CaT), where every constraint is independent of the skill command so that behavioral differences arise only from the video reward. A symmetry loss on the policy is added because the camera sees the robot from one side only. The argument therefore rests on a division of labor: the video classifier supplies what each skill means, and the embodiment-agnostic constraints supply what is physically possible for the robot.","core_discovery":"The paper establishes that a video action classifier trained purely on animals in natural habitats can act as a valid scalar reward for robotic locomotion, bridging what the authors call the extreme gap in domain and embodiment between animals and robots. A Uniformer video model is fine-tuned on a single-label subset of the Animal Kingdom dataset covering keeping still, walking, running, and jumping, then frozen. During policy training in a physics simulator, a third-person camera records 8-frame clips of the robot and the classifier's probability for the commanded skill is used as the reward, with zero reward on non-rendering steps and per-skill reward normalization. Task-agnostic constraints (joint limits, torques, foot air time, base orientation) enforce physical plausibility and sim-to-real transferability, and a symmetry loss compensates for the single camera view. The authors report that distinct, recognizable behaviors emerge for each skill command and that the policy deploys on a physical Solo-12 robot, with walking and running appearing as trotting gaits and jumping as rhythmic in-place pumping with flight phases; they also concede that running never produces a proper flying phase and that the behaviors lag the state of the art in learning-based locomotion.","pith_inferences":["Editorial inference: the method turns reward engineering into a recognition problem, so the quality ceiling is set by how well the frozen classifier generalizes out of distribution; improvements in video understanding should translate directly into better locomotion skills without changing the RL loop.","Editorial inference: the two-leg walking-in-place failure reported for the walking skill shows the reward can be gamed by visual artifacts; a direct test would measure how strongly the classifier score correlates with forward velocity and foot contact across rollouts, and a fix would add multi-view or motion-focused scoring.","Editorial inference: the same scheme could point at other video corpora — human sports footage, finer-grained animal behavior labels, or egocentric clips — to produce rewards for skills that are hard to specify by hand, whenever the visual gap between source and robot is comparable to the one tested here.","Editorial inference: the sensitivity to dataset size (skills degrade when training drops below 50% of the 8,791 videos) suggests the reward's robustness scales with video diversity rather than with the RL algorithm, making internet-scale data the main lever for future work."],"forward_implications":["A video classifier trained on animal footage can reward a quadruped policy, so the same recipe should extend to other skills and other legged embodiments whenever the classifier can recognize the skill across morphologies.","The no-curating ablation shows the reward signal must be tailored to single labels: training the classifier on the full multi-label dataset with binary cross-entropy prevents walking and running from emerging at all.","The learned running is only a faster trot without an aerial phase, and walking and running look similar on the real robot; the paper attributes this gap to conventional video classification and standard on-policy RL rather than to the overall approach.","Sim-to-real transfer succeeds directly, without domain randomization, a result the authors attribute to the task-agnostic constraints — especially the foot air-time constraint — rather than to the reward function itself.","Because rewards fire every five steps from one camera, the symmetry loss and a camera angle that shows the full body are necessary for walking and running to emerge; extreme camera angles degrade or destroy individual skills."],"supporting_citations":[{"why":"Supplies the Animal Kingdom dataset of labeled wild-animal videos used to train the video classifier that becomes the reward.","marker":"[11]"},{"why":"Provides the Uniformer video-transformer architecture, pretrained on Kinetics-400, that the classifier is fine-tuned from; its out-of-distribution performance determines reward quality.","marker":"[91]"},{"why":"Model soups (weight averaging of fine-tuned classifiers) improve generalization to robot videos; the no-soup ablation shows this supports skill emergence.","marker":"[93]"},{"why":"Random convolution augmentations applied during classifier training to improve out-of-distribution generalization to robot renders.","marker":"[92]"},{"why":"CaT (Constraints as Terminations), the constrained-RL mechanism that enforces the task-agnostic constraints during policy training.","marker":"[20]"},{"why":"PPO, the on-policy RL optimizer used to train the multi-skill policy against the video-based reward.","marker":"[18]"},{"why":"The massively parallel simulation training paradigm the paper follows to learn locomotion policies efficiently.","marker":"[17]"},{"why":"Isaac Gym, the physics simulator where the robot policy is trained and the third-person reward videos are rendered.","marker":"[16]"},{"why":"The symmetry loss on the policy that compensates for the single-camera reward observation during RL.","marker":"[95]"},{"why":"The open torque-controlled Solo-12 quadruped robot used for the real-world deployment of the learned policy.","marker":"[97]"}],"fun_headline_variants":["Robot learns to walk from nature documentaries","Wild animal videos teach robot to walk and jump","No hand-coded rewards: robot learns from animal clips","Video classifier turns wild animal clips into robot rewards","Robot masters walking and jumping by watching wildlife"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that the video classifier, trained only on wild-animal clips, scores the simulated and real robot's movements for the right reason — recognizing the skill in the motion — rather than because of background, camera angle, body pose, or incidental visual artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Robot learns to walk from nature documentaries","Wild animal videos teach robot to walk and jump","No hand-coded rewards: robot learns from animal clips","Video classifier turns wild animal clips into robot rewards","Robot masters walking and jumping by watching wildlife"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000318,"raw_usage":{"total_tokens":1802,"prompt_tokens":956,"completion_tokens":846,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":790}},"tokens_in":572,"tokens_out":846,"duration_ms":7977,"temperature":1.0,"reasoning_tokens":790,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:33:40.797698+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed the trained classifier robot clips with the body masked or blurred while the background and camera motion are preserved: if the 'Walking' or 'Running' probability stays high, the reward is tracking visual confounds rather than the skill. A direct correlation measurement between the classifier score and ground-truth forward velocity or foot air time across policy rollouts would likewise reveal how much of the reward is skill-grounded, and the two-leg walking-in-place failure reported in the paper is a concrete case where the score stayed high while the skill was absent.","supporting_citations":[{"cited_title":"Wortsman, G","cited_arxiv_id":null,"evidence_quote":"Model soups (weight averaging of fine-tuned classifiers) improve generalization to robot videos; the no-soup ablation shows this supports skill emergence."},{"cited_title":"Grimminger, A","cited_arxiv_id":null,"evidence_quote":"The open torque-controlled Solo-12 quadruped robot used for the real-world deployment of the learned policy."}],"review_version":1}