{"id":"8668493e-21ec-4954-8a97-1f51ed16ace2","arxiv_id":"2508.12252","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A force-sensing arm acts as teacher for a small humanoid, enabling 20-minute real-world walking speed adaptation and 15-minute swing-up learning from scratch.","lead":"This paper builds a 'robot trains robot' setup: a force-sensing robot arm holds, steadies, measures, and resets a small humanoid while a reinforcement learning policy is retrained in the real world. The authors report that the setup doubles the humanoid's walking speed after 20 minutes of training and teaches it a swing-up motion from scratch in 15 minutes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The speed-doubling and swing-up results are measured through the teacher's own force/treadmill feedback loop, with no independent speed or swing-angle measurement; the proxy reward in Eq. 9 can be gamed by tether pulling.","rationale":"The reader identifies the same load-bearing assumption: the walking reward is computed from a treadmill speed that is itself a control output of the RTR teacher (Eq. 9), and the paper provides no independent verification that this quantity equals the humanoid's actual walking speed. I agree that this is the most serious weakness in the central claim. The concern is not a disagreement with the field's consensus; it is a correctness risk about the reward and measurement channel. The paper's ablations and the RMA comparison are real evidence that some behavior changes, but they do not establish that the measured quantity is the student's walking speed or swing amplitude. Appendix C.5 being empty and the absence of any numerical speed-doubling result make the claim impossible to audit from the text. Because the reader's conditional verdict already reflects this uncertainty, I do not recommend changing the verdict; however, the concrete test above should be run before the speed-doubling claim is accepted at face value.","tokens_in":16175,"tokens_out":7523,"duration_ms":85190,"concrete_test":"Run a walking evaluation with the Eq. 9 feedback disabled (k1^p = k2^p = 0, belt fixed at 0.15 m/s) while measuring torso forward velocity with an external motion-capture system, and report the trained policy's speed-tracking error and its maximum trackable belt speed before and after fine-tuning. If the RTR policy does not produce forward torso velocity matching the belt speed, or if its advantage over Freeze pi(s,z) disappears when the F_x/psi feedback terms are zeroed, then the treadmill-speed reward in Section 3.2 was the confound. For the swing-up task, the analogous check is to record rope angle with an encoder and compare it with the F/T-based force amplitude used in the reward.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Both headline demonstrations are validated through the teacher's own measurement loop, and no independent ground-truth measure of the learned skill is reported. For walking, the reward is r = exp(-sigma * (v - v_target)^2) (Section 3.2) with v 'approximated' by treadmill speed, and Appendix C.3 defines that treadmill speed as v = v_base + k1^p * F_x + k2^p * psi (Eq. 9), where F_x is the force the humanoid exerts on the elastic tether and psi is torso pitch. A policy can therefore inflate its own reward by pulling on the tether (positive F_x) or by adopting a posture that drives psi, without increasing actual gait speed. The paper's 'doubles the zero-shot walking speed' headline is never backed by a measured walking speed, a tracking-error curve, or a before/after maximum-speed number; Table 1 only reports stability metrics at one belt speed, and Appendix C.5, which would contain the real-world protocol, is empty. For swing-up, the reward is the FFT amplitude of the same force sensor that the arm drives during phase-aligned helping (x_t = x_0 + A_arm * cos(theta_t)), and no rope-angle or swing-height measurement is reported. The central demonstrations are thus consistent with optimizing the teacher's sensor loop rather than the student's actual walking or swinging.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Robot-Trains-Robot (RTR), a system in which a force-sensing robot arm acts as a teacher that supports, guides, rewards, perturbs, resets, and schedules a small humanoid student during real-world reinforcement learning. The learning method has three stages: training a dynamics-conditioned policy with FiLM modulation across domain-randomized simulation environments, optimizing a universal dynamics latent in simulation, and then fine-tuning only that latent in the real world with PPO. The paper reports two hardware demonstrations on the ToddlerBot platform: fine-tuning a walking policy for treadmill speed tracking, claimed to double the zero-shot walking speed with 20 minutes of real-world training, and learning a swing-up behavior from scratch in 15 minutes. Ablations cover arm compliance, arm height scheduling, latent versus policy fine-tuning, FiLM learning rate, and a comparison with RMA.","tokens_in":16472,"tokens_out":6201,"duration_ms":64165,"significance":"If the quantitative claims are supportable, RTR is a significant systems contribution: it is one of the few demonstrations of autonomous real-world humanoid RL with minimal human intervention, and the dynamics-latent fine-tuning pipeline is a sensible way to keep real-world updates low-dimensional. Strengths include three-seed ablations for the main comparisons, explicit baselines (fine-tuning the base policy, fine-tuning a residual policy, fixed-arm and fixed-latent variants, and RMA), a FiLM learning-rate ablation with zero-latent evaluation, and the use of an open-source, low-cost humanoid platform. The main unresolved issue is measurement validity: the headline results are evaluated through the teacher's own force and treadmill feedback loop rather than through independent measurements of the learned skill.","major_comments":[{"comment":"The headline claim that RTR 'doubles the zero-shot walking speed' is not backed by an independent measurement of walking speed. The reward in Eq. (3) is computed from v, which the paper says is 'approximated' by the treadmill speed, and Appendix C.3 defines that treadmill speed as v = v_base + k1^p Fx + k2^p psi, where Fx is the force the humanoid exerts on the elastic tether and psi is torso pitch. A policy can therefore increase its own reward by pulling on the tether or adopting a posture that drives psi without increasing actual gait speed. No before-and-after speed value, speed-tracking error curve, or maximum-speed number appears in Section 4.1 or Table 1; Table 1 reports only stability metrics at a single belt speed of 0.15 m/s. Please provide an independent measurement of the robot's gait speed (for example, motion-capture torso velocity, foot-contact-based speed, or camera tracking) and report the actual pre/post speed values that support the 'doubles' claim.","section":"Section 3.2, Eq. (9), and Section 4.1"},{"comment":"The swing-up result is also measured through the teacher's own sensor loop. The reward in Eq. (4) is the FFT amplitude of the force sensor at the dominant frequency, while the teacher arm's helping motion is phase-aligned to the same estimated swing phase, with x_t = x0 + A_arm cos(theta_t). The paper reports no independent rope-angle, IMU-based tilt, or vision-based swing-height measurement, so the learned behavior could be an interaction with the arm rather than a true pendulum swing-up. Please add an independent swing-angle or swing-height trace, or otherwise validate quantitatively that the force amplitude tracks the actual swing amplitude.","section":"Section 3.3, Eq. (4), and Figure 5"},{"comment":"Appendix C.5, labeled 'Real-world Learning Details', is an empty heading with no content. This is exactly the section that would document the real-world protocol for both tasks, including reset procedures, batch timing, evaluation protocol, and the measurements behind the 20-minute and 15-minute claims. Its absence makes the headline experiments difficult to reproduce and prevents the reader from assessing the proxy-reward concerns above. The section should be filled in or its content should be integrated into the main experimental details.","section":"Appendix C.5"},{"comment":"The walking ablations are evaluated with the same proxy reward that is being optimized. The y-axis 'linear velocity tracking rewards' is computed from the Eq. (9) treadmill speed, so the training curves conflate genuine policy improvement with exploitation of the force and pitch feedback loop. The 'better data efficiency' conclusion should be re-stated in terms of the measured walking outcome once an independent speed measurement is added; as written, the conclusion is about proxy reward rather than gait speed.","section":"Figure 4 and Section 4.1"}],"minor_comments":[{"comment":"The heading 'Rapad Motor Adaptation' should be 'Rapid Motor Adaptation'.","section":"Appendix D"},{"comment":"The text refers to 'appliance control'; from the context this appears to be a typo for 'admittance control'.","section":"Appendix C.2"},{"comment":"The sentence 'we apply an larger entropy coefficient' should be 'we apply a larger entropy coefficient'.","section":"Section 4.2"},{"comment":"The expression 'theta_0 = 30 degrees' uses an unusual degree symbol; use the standard notation for consistency.","section":"Section 3.3"},{"comment":"Reference [10], a market research report, is an unusual source for the claim about industrial arm payload capacity; a technical datasheet or manufacturer specification would be more appropriate.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The main concern is measurement validity, not the quality of the hardware effort or the novelty of the system. The manuscript is currently in a state where the central quantitative claims are plausible but unverified by independent sensors, and Appendix C.5 is empty. I recommend major revision rather than rejection because the central framework is defensible and the missing measurements appear obtainable from the existing setup. Please ask the authors for the actual walking-speed values behind the doubling claim, an independent swing-up measurement, and a completed Appendix C.5."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know before reading. First, RTR is a credible engineering contribution: a UR5 arm with F/T sensing, compliant support, auto-reset, and a curriculum is exactly the kind of scaffolding that could make long-horizon real-world humanoid RL practical, and the paper demonstrates it on two tasks. Second, the two headline numbers—20-minute adaptation, 15-minute swing-up, \"doubles zero-shot walking speed\"—are not actually quantified in the experiments. The paper shows reward curves and stability metrics at one treadmill speed; the doubling claim appears in the abstract and intro, but no before/after speed value or tracking-error curve appears anywhere in the evaluation.\n\nWhat is genuinely new is the package. The teacher-arm idea extends assistive-force curricula, but RTR adds a real-time force-based reward channel, failure detection, automatic resets, and a three-stage latent fine-tuning pipeline (FiLM-conditioned dynamics policy, universal latent, real-world latent update). The ablations are decent: three seeds, explicit baselines, and a fairly detailed comparison with RMA in simulation and on hardware. The paper also openly lists limitations, including task-specific curriculum design and missing ground-reaction force sensing. That honesty deserves credit.\n\nThe soft spots are real but not necessarily fatal. The most important one: the walking reward is computed from treadmill speed, and the treadmill speed is itself a feedback output of the teacher's controller (Eq. 9), v = vbase + k1*Fx + k2*psi, with Fx the force the humanoid exerts on the elastic tether. A policy that learns to pull the tether or bias its posture can inflate the reward without walking faster. The paper states it approximates walking speed with treadmill speed, but never validates that proxy against an independent gait-speed measurement. Similarly, the swing-up reward is the FFT amplitude of the force sensor that the arm itself drives during phase-aligned helping; no rope-angle or swing-height measurement is reported. So the current experiments are consistent with optimizing the teacher's sensor loop rather than the student's actual walking or swinging. I do not think this destroys the qualitative claim—something is clearly being learned—but it does mean the quantitative headlines overstate what is shown.\n\nAnother concrete defect: Appendix C.5, titled \"Real-world Learning Details,\" is an empty heading. That is exactly the section that could have contained the missing protocol. Minor point: no code or data release, which makes independent verification harder. The citation pattern is fine; the related work is appropriately broad.\n\nWho should read it: anyone working on real-world humanoid RL or teacher-student training systems. It deserves a serious referee, not a desk reject, but the referee should ask for independent measurements or, failing that, scaled-back claims. I would not cite the speed-doubling number in my own work; I might cite the system if I needed a reference for the teacher-arm idea.","headline":"RTR is a credible and useful hardware system, but the headline speed-doubling claim is not backed by a direct speed measurement and the walking reward loop is coupled to the teacher's own force/treadmill feedback, so the quantitative claims need careful handling.","tokens_in":17068,"tokens_out":2598,"would_cite":false,"duration_ms":28430,"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":"Robot-Trains-Robot claims a force-sensing robot arm can teach a humanoid to walk and to swing up, using only minutes of real-world training.","keywords":["humanoid robots","real-world reinforcement learning","sim-to-real transfer","dynamics latent optimization","teacher-student robot learning","compliant robot arm","walking speed tracking","swing-up learning"],"falsifier":"Measure the humanoid's true forward velocity independently with motion capture or with an overground trial while RTR trains; if the treadmill speed stays high while true walking speed does not rise, or if the learned policy fails to transfer to a treadmill without the tether feedback loop, the central speed-doubling claim would be falsified.","tokens_in":15943,"feed_emoji":"🤖","tokens_out":7414,"duration_ms":67012,"temperature":0.7,"pith_summary":"The paper introduces Robot-Trains-Robot (RTR), a system in which a robot arm with force sensing acts as teacher for a humanoid student: it supports the student safely, supplies reward signals that would otherwise be unavailable, runs an automatic curriculum, detects failures, and resets the robot without human help. The central claim is that this setup makes real-world reinforcement learning practical on humanoids. In the walking task, fine-tuning a single dynamics-encoded latent variable with RTR doubles the zero-shot walking speed after about 20 minutes of real-world training. In the swing-up task, the humanoid learns a periodic swing-up motion from scratch within about 15 minutes of real-world interaction. If correct, this is evidence that the missing ingredient for real-world humanoid learning is not the RL algorithm alone but a physical teacher that closes the safety, reward, and reset loops.","feed_headline":"Robot-arm teacher doubles humanoid walking speed in 20 minutes","feed_subtitle":"A force-sensing arm supplies safety, rewards, and resets, letting a humanoid learn real-world skills from scratch.","key_machinery":"The load-bearing object is the dynamics latent $\\mathbf{z}$, a single vector that encodes environment physics. In Stage 1, an encoder $f_\\phi$ maps randomized physics parameters $\\mu^{(i)}$ to $\\mathbf{z}^{(i)}$, and Feature-wise Linear Modulation (FiLM) layers modulate the actor's hidden states by scaling and shifting them, so the policy becomes dynamics-conditioned. Stage 2 optimizes a universal latent $\\tilde{\\mathbf{z}}$ shared across all simulated environments to give a reliable start. Stage 3 freezes the actor and FiLM parameters and fine-tunes $\\tilde{\\mathbf{z}}$ in the real world with PPO, so real-world adaptation is reduced to optimizing one low-dimensional latent instead of the full network. The hardware counterpart is the admittance-controlled teacher arm with a force-torque sensor, plus an optional treadmill whose speed is PD-controlled from tether force and torso pitch; these provide safety, the speed proxy used as reward, and automatic reset.","core_discovery":"The paper's central discovery is that a teacher arm with force-torque sensing can supply everything a humanoid policy needs to learn in the real world: safe exploration via compliant support, reward via measured interaction, curriculum via scheduled support height and helping or perturbing arm motions, and automation via failure detection and resets. For sim-to-real adaptation, the paper proposes a three-stage pipeline: a dynamics-aware policy is pretrained in domain-randomized simulation with a latent vector encoded from physics parameters and injected through FiLM layers; a universal latent is optimized across all simulated environments; then only that latent is fine-tuned in the real world while the actor and FiLM layers stay frozen. This one-latent fine-tuning is what yields the reported speed-tracking improvement. For learning from scratch, the same physical infrastructure provides a helping and perturbing arm schedule and offline critic pretraining, letting the humanoid discover a swing-up motion directly in the real world.","pith_inferences":["Editorial inference: the walking reward loop couples treadmill speed to tether force, so a policy that learns to pull the tether could raise its own reward without walking faster; an overground transfer test would separate true walking from tether exploitation.","Editorial inference: the helping and perturbing arm schedule is a physical curriculum that pumps energy into the system, suggesting the same mechanism could train other underactuated or unstable robots whenever an external agent can safely add or remove energy.","Editorial inference: if the universal-latent initialization generalizes, the same three-stage recipe could serve as a sim-to-real warm start for other legged robots, with the latent encoding terrain, friction, or payload rather than only humanoid dynamics."],"forward_implications":["If the central claim holds, real-world fine-tuning of a humanoid walking policy can be reduced to optimizing a single latent vector, making adaptation far more data-efficient than full-network fine-tuning.","Real-world learning from scratch becomes feasible for tasks that are difficult to simulate, such as cable-suspended swing-up, because the teacher arm provides safety and a physical curriculum without a scripted policy.","The same infrastructure can double as a data-collection, failure-detection, and reset system, cutting human supervision to the point of long, mostly unattended training sessions.","The system is platform-agnostic in principle: any robot arm or crane with force sensing could serve as teacher for larger humanoids whenever payload and workspace allow."],"supporting_citations":[{"why":"Supplies the open-source ToddlerBot humanoid platform used for all real-world experiments.","marker":"[9]"},{"why":"PPO is the reinforcement learning algorithm used in all three training stages and in the real-world updates.","marker":"[8]"},{"why":"Domain randomization is the basis for Stage 1 dynamics-conditioned pretraining.","marker":"[6]"},{"why":"FiLM layers are the mechanism that injects the dynamics latent into the actor network.","marker":"[7]"},{"why":"RMA is the closest latent-based sim-to-real adaptation baseline and is compared in simulation and real world.","marker":"[30]"},{"why":"Its online-offline three-stage scheme motivates the swing-up critic pretraining pipeline.","marker":"[35]"},{"why":"Provides the admittance control concept used for compliant arm guidance.","marker":"[53]"},{"why":"Supports the claim that industrial arms with sufficient payload could extend the teacher paradigm to full-scale humanoids.","marker":"[10]"}],"fun_headline_variants":["Robot-arm tutor lets humanoid learn real-world skills","Humanoid learns swing-up from scratch with a robot teacher","One latent variable fine-tunes humanoid control in reality","Robot trains robot: arm-guided humanoid learning in minutes","Teacher arm provides safety and resets for real-time humanoid learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The walking result assumes that the treadmill speed, which is itself computed from the humanoid's pull on the elastic tether and its torso pitch, faithfully measures how fast the humanoid is actually walking; if the policy instead learns to exploit the tether to raise its own reward, the speed-tracking result would not be about walking speed.","fun_headline_variants_meta":{"raw":{"variants":["Robot-arm tutor lets humanoid learn real-world skills","Humanoid learns swing-up from scratch with a robot teacher","One latent variable fine-tunes humanoid control in reality","Robot trains robot: arm-guided humanoid learning in minutes","Teacher arm provides safety and resets for real-time humanoid learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000453,"raw_usage":{"total_tokens":2277,"prompt_tokens":945,"completion_tokens":1332,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":1250}},"tokens_in":561,"tokens_out":1332,"duration_ms":15149,"temperature":1.0,"reasoning_tokens":1250,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:25:00.434876+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the humanoid's true forward velocity independently with motion capture or with an overground trial while RTR trains; if the treadmill speed stays high while true walking speed does not rise, or if the learned policy fails to transfer to a treadmill without the tether feedback loop, the central speed-doubling claim would be falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the open-source ToddlerBot humanoid platform used for all real-world experiments."},{"cited_title":"Kumar, Z","cited_arxiv_id":null,"evidence_quote":"RMA is the closest latent-based sim-to-real adaptation baseline and is compared in simulation and real world."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Its online-offline three-stage scheme motivates the swing-up critic pretraining pipeline."},{"cited_title":"Industrial robotic arm market size, share, growth trends, regional share, competitive intelligence, forecast report 2025–2037, March 2025","cited_arxiv_id":null,"evidence_quote":"Supports the claim that industrial arms with sufficient payload could extend the teacher paradigm to full-scale humanoids."}],"review_version":2}