{"id":"0491d77c-c4e9-4751-8b7b-773a6d4c40b3","arxiv_id":"2508.02194","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"Constrained reinforcement learning with domain randomization transferred a balance and velocity control policy from simulation to a real point-foot, armless bipedal robot.","lead":"Using reinforcement learning with safety constraints, the authors trained the two-legged robot Bolt, which has point feet and no arms, to walk and keep its balance in simulation, then transferred the trained policy to the real robot. It matters to robotics because point-foot bipeds are hard to stabilize, and a working sim-to-real pipeline would make such simple, low-cost robots more practical.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The sim-to-real transfer claim is unsupported without evidence of hardware deployment; the abstract does not state whether experiments used the physical Bolt robot.","rationale":"The reader's weakest assumption concerned simulator fidelity and domain randomization ranges. My concern is more direct: the abstract's own wording may not even claim a demonstrated hardware result, only a methodology that 'enables' transfer. The absence of explicit hardware evidence is the most load-bearing issue because the paper's stated contribution is sim-to-real transfer. I agree with the UNVERDICTED outcome, but my emphasis is on missing evidence of physical deployment rather than on simulator fidelity. A single well-placed sentence in the full text could settle this. Since the full text is unavailable, the verdict remains unverdictable, so I recommend UNCHANGED.","tokens_in":673,"tokens_out":1387,"duration_ms":16206,"concrete_test":"Obtain the full text and locate the experimental protocol. Specifically, search for 'hardware,' 'physical Bolt,' 'real-robot,' or 'deployment.' If the results section reports only simulated rollouts and no physical robot trials, the transfer claim is not supported. If hardware trials are reported, verify that they include disturbance tests matching the abstract's slip and push scenarios.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the methodology 'enables sim-to-real transfer.' For this to be true, the policy trained in simulation must be demonstrated on the physical Bolt robot. The abstract mentions qualitative and quantitative experiments but does not explicitly state that any experiments were performed on real hardware; they could all be simulated, with 'sim-to-real' describing only a design intention. The claim of advancing robust control strategies for point-foot bipeds rests on successful transfer. If no hardware trials were conducted, the claim reduces to a hypothesis rather than a demonstrated result. This is a load-bearing concern because the paper's contribution is the transfer itself, not just a simulation study.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a constrained reinforcement learning approach for the point-foot Bolt bipedal robot, combining Constraints-as-Terminations with domain randomization to enable sim-to-real transfer. The abstract describes qualitative and quantitative experiments evaluating balance maintenance, velocity control, and responses to slip and push disturbances, and it reports analysis of cost of transport and ground reaction force. No numerical results, experimental details, or explicit hardware outcomes appear in the abstract.","tokens_in":768,"tokens_out":2114,"duration_ms":25095,"significance":"If the claimed sim-to-real transfer is actually demonstrated, this work would contribute to robust control of underactuated point-foot bipedal robots, a challenging and relevant class of systems. The combination of constrained RL with termination-based constraints and domain randomization is a reasonable methodological direction. However, as presented in the abstract, there is no quantitative evidence, no comparison to baselines or prior work, no ablations, and no explicit statement that any experiment was performed on the physical Bolt robot. The significance of the contribution therefore cannot be assessed from the abstract alone.","major_comments":[{"comment":"The central claim that the methodology 'enables sim-to-real transfer' is not supported anywhere in the abstract, because the text never states that experiments were performed on the physical Bolt robot. If the described qualitative and quantitative experiments are all simulated, then the transfer claim is only a design intention, not a demonstrated result. The full paper must explicitly state which experiments involve real hardware and report those hardware results.","section":"Abstract"},{"comment":"The abstract reports no numerical results, error bars, or comparisons, so the claims about balance maintenance, velocity control, and disturbance handling cannot be verified. Even for an abstract, at least one representative quantitative metric (e.g., mean tracking error, success rate under pushes, or cost of transport) should be stated so that the reader can judge whether the claims have empirical support.","section":"Abstract"},{"comment":"The evaluation metrics listed (balance, velocity tracking, cost of transport, and ground reaction force) plausibly align with the reward function and constraint thresholds used during training, but the abstract gives no information about how these metrics are defined or whether they are distinct from the training objective. Without such a distinction, the evaluation risks being partially circular; the full paper should specify the reward terms, constraint thresholds, and evaluation protocol separately.","section":"Abstract"}],"minor_comments":[{"comment":"The phrase 'analyze autonomy through metrics like the cost of transport and ground reaction force' conflates energy efficiency and foot loading with autonomy; the paper should either define what is meant by autonomy or use a more precise term.","section":"Abstract"},{"comment":"The terms 'Constraints-as-Terminations' and 'domain randomization' are introduced without citations or definitions; the full paper should reference the prior work from which these techniques are taken.","section":"Abstract"},{"comment":"The abstract says 'a series of qualitative and quantitative experiments' but gives no indication of the number of trials, the robot configuration, or the statistical treatment of the results; this information should appear in the full text and ideally be summarized in the abstract.","section":"Abstract"}],"recommendation":"uncertain","confidential_remarks":"This review is based only on the abstract because the full text was not available. The main uncertainty is whether the full manuscript contains actual hardware experiments on the Bolt robot; if it does not, the sim-to-real transfer claim must be substantially weakened. I recommend that the editor obtain the full text before making a decision, and pay particular attention to the experimental section and to whether the evaluation metrics are independent of the training objective."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is an abstract-only submission, so the only honest verdict is provisional. The paper claims sim-to-real transfer for the Bolt point-foot biped using Constraints-as-Terminations and domain randomization. That combination on that hardware is a genuinely useful thing to try, and the abstract lays out sensible evaluation axes (balance, velocity tracking, slip/push response, cost of transport, ground reaction force). If the full paper shows the policy running on the physical robot with numbers and comparisons, it is worth reading.\n\nThe soft spot is exactly where the stress-test note points: the abstract never explicitly says the experiments were on real hardware. \"Sim-to-real transfer\" can mean \"we designed the training to transfer\" or \"we ran it on the robot.\" For this community, only the second reading supports the central claim. The abstract says \"qualitative and quantitative experiments\" but that could all be in simulation. That is a load-bearing omission, not a stylistic one. Also missing from the abstract: any numbers, error bars, baselines, or details on randomization ranges and constraint thresholds. Without those, the central claim is a teaser. The abstract also cites no prior work, so novelty cannot be assessed against the literature.\n\nI don't want to manufacture a fatal flaw. The full text may well include hardware trials. But the abstract as written is not enough to verify anything, and the stress-test concern is fair. If the full paper has real-robot results, the missing details are just a bad abstract; if not, the paper needs to be reframed as a simulation study with a transfer recipe, which is much weaker.\n\nWho is this for? People working in legged locomotion, sim-to-real for underactuated systems, and constrained RL. A serious referee should see the full text before any verdict. I would not accept the abstract as evidence, but I would send the paper out for review if there is any sign of hardware experiments. The editor should ask the authors directly whether the Bolt trials were physical.\n\nRecommendation: send to peer review, because the problem is hard and the approach is plausible; condition the outcome on explicit hardware evidence and quantitative comparisons.","headline":"Abstract-only paper with a plausible sim-to-real claim for a hard biped; the missing explicit hardware confirmation is the key issue to resolve.","tokens_in":1297,"tokens_out":1997,"would_cite":false,"duration_ms":23176,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Constrained RL transfers a point-foot biped to real hardware","keywords":["constrained reinforcement learning","sim-to-real transfer","bipedal locomotion","point-foot robot","Constraints-as-Terminations","domain randomization","Bolt robot","underactuated control"],"falsifier":"Deploy the trained policy on the physical Bolt and subject it to the same push and slip disturbances used in simulation; if the robot falls or fails to track commanded velocities within the randomized parameter ranges, the claimed sim-to-real transfer fails.","tokens_in":507,"feed_emoji":"🤖","tokens_out":3790,"duration_ms":43403,"temperature":0.7,"pith_summary":"The paper sets out to show that a point-foot bipedal robot, the Bolt, can be stabilized and controlled by a reinforcement learning policy despite being underactuated, having no arms, and having limited foot actuation. The proposed methodology combines Constraints-as-Terminations with domain randomization so that a policy trained entirely in simulation is robust enough to transfer to the real robot. If the claim is right, the same recipe gives a practical route to deploy learned controllers on unstable legged hardware without explicit dynamics modeling. The paper supports the claim with qualitative and quantitative experiments on balance maintenance, velocity tracking, and responses to slip and push disturbances, plus autonomy metrics such as cost of transport and ground reaction force.","feed_headline":"Constrained RL transfers a point-foot biped to real hardware","feed_subtitle":"Bolt's underactuated, armless point-foot design makes balance a hard test for sim-to-real control.","key_machinery":"Constraints-as-Terminations (CaT): a reinforcement-learning training rule in which any violation of a constraint (for example, a joint limit or a fall) immediately ends the episode, so the agent learns to avoid those states without needing a hand-tuned penalty schedule. Domain randomization: randomly varying simulator parameters such as masses, friction coefficients, and actuation gains across training episodes, forcing the policy to behave well across a distribution of dynamics. Together these two mechanisms carry the sim-to-real transfer claim: CaT shapes the behavior to stay within safe limits, while domain randomization covers the gap between simulation and the physical Bolt.","core_discovery":"The central claim is that constrained reinforcement learning, specifically Constraints-as-Terminations paired with domain randomization, produces a control policy for the point-foot Bolt that transfers from simulation to the real robot. The robot is a hard case because it is underactuated, has no arms to rebalance, and has limited actuation at the feet. The paper argues that treating constraint violations as episode terminations, rather than as shaped penalties, steers the policy away from unsafe states, while domain randomization across training episodes makes the behavior robust to sim-to-real mismatch. In the authors' telling, the resulting policy maintains balance and tracks velocity commands on hardware, and withstands slip and push disturbances.","pith_inferences":["Beyond the paper: the same constraint-as-termination recipe could be applied to other underactuated platforms, such as wheeled-legged robots, since it only requires a simulator and a constraint detector.","Beyond the paper: systematically varying the randomization ranges and measuring real-world fall rate would quantify how much robustness margin the policy actually has, a quantity the paper does not report.","Beyond the paper: comparing CaT against a shaped constraint penalty head-to-head in simulation would isolate whether terminations alone, rather than the reward structure, drive the transfer success."],"forward_implications":["The trained policy can keep the physical Bolt upright while tracking velocity commands, despite the robot's lack of arms and underactuated point feet.","The policy should remain stable under slip and push disturbances that fall within the domain randomization ranges used in training.","Autonomy metrics such as cost of transport and ground reaction force should reflect economical, non-spastic gaits rather than corrective micro-steps.","The methodology gives a general recipe for applying constrained RL to other unstable legged robots without adding hardware safety filters."],"supporting_citations":[],"fun_headline_variants":["RL with termination constraints stabilizes point-foot Bolt in real tests","Point-foot biped Bolt balances via constrained sim-to-real RL","Underactuated Bolt walks using constrained RL and randomization","Constrained RL transfers point-foot biped control to real Bolt"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that the simulator, with its domain randomization ranges, represents the real Bolt closely enough that a policy learned there will work on the hardware.","fun_headline_variants_meta":{"raw":{"variants":["RL with termination constraints stabilizes point-foot Bolt in real tests","Point-foot biped Bolt balances via constrained sim-to-real RL","Underactuated Bolt walks using constrained RL and randomization","Constrained RL transfers point-foot biped control to real Bolt"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1351,"prompt_tokens":804,"completion_tokens":547,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":420,"completion_tokens_details":{"reasoning_tokens":476}},"tokens_in":420,"tokens_out":547,"duration_ms":6954,"temperature":1.0,"reasoning_tokens":476,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:05:25.172073+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy the trained policy on the physical Bolt and subject it to the same push and slip disturbances used in simulation; if the robot falls or fails to track commanded velocities within the randomized parameter ranges, the claimed sim-to-real transfer fails.","supporting_citations":[],"review_version":1}