{"id":"011fc0e9-a1b2-4555-be6a-39c5038512ad","arxiv_id":"2501.05204","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A bipedal entertainment robot with backward-bending knees and antennas is controlled by multiple reinforcement-learning policies, an animation blending engine, and a two-joystick puppeteering interface, demonstrated in public shows.","lead":"This paper presents a new bipedal robot character from Disney Research and a reinforcement-learning control pipeline that lets a human operator puppeteer it in real time. The work shows how artist-driven animation, mechanical design, and robust legged locomotion can be combined into a single believable character.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sim-to-real transfer is the load-bearing assumption; the test-bench actuator model may not capture in-situ power/thermal coupling, and the deployment evidence does not directly bound the sim-to-real gap.","rationale":"We agree with the reader that sim-to-real transfer is the load-bearing assumption. Our pass refines this to the in-situ validity of the test-bench actuator model: the robot's shared power bus and thermal environment can create torque deficits that the per-actuator identification and domain randomization do not cover. The deployment evidence is meaningful but aggregate; it does not bound the gap under sustained high-power episodic motions, which are precisely the expressive behaviors the central claim highlights. We therefore keep the CONDITIONAL verdict. The proposed concrete test would either validate the margin or reveal a training envelope mismatch, either of which directly speaks to whether the policies are robust as claimed. We considered other possible concerns (policy-switch off-distribution actions, lack of user study) and found them either secondary or not correctness risks.","tokens_in":18165,"tokens_out":10269,"duration_ms":105957,"concrete_test":"Run a controlled high-power episode (e.g., 10 consecutive jump motions) on hardware while logging battery voltage, per-actuator current, and torque. Replay the same command sequence in Isaac Gym with the identified actuator model and randomized parameters. Compare the per-joint torque-tracking error and the time to reach torque limits. If hardware torque degrades by more than the randomized range (e.g., a 20% drop in effective tau_max) during the sequence, or if the simulated and hardware contact timings diverge by more than 0.1 s, the sim-to-real margin is too thin to support the central claim. Alternatively, identify actuator parameters in situ from these logs and check whether they fall inside the randomized ranges; if not, the deployment evidence is insufficient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Central to the paper's claim is that Isaac Gym-trained policies transfer to the physical robot without fine-tuning (Sec. V-E, App. B). The sim actuator model is identified on an isolated test bench; on the 15.4 kg robot, the same actuators share a battery, so high-power episodic motions (e.g., the jump shown in Fig. 8) cause voltage sag and thermal drift that reduce torque at the joint. The domain randomization only covers isolated-actuator parameter ranges, not coupled power/thermal effects. The 10 h of no-fall runtime is strong evidence but aggregate; it does not quantify the margin or test sustained high-power episodes. If in-situ torque-speed curves exceed the randomized envelope, the expressive-motion robustness claim is not substantiated. This is a boundary-condition concern, not an internal inconsistency.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a complete workflow for creating a bipedal robot character: mechanical design driven by artistic intent, simulation-based training of multiple reinforcement-learning policies for standing, walking, and episodic motions, a runtime animation engine that blends background, triggered, and joystick-driven animation layers, and a two-joystick puppeteering interface. The system is evaluated through joint-tracking errors (Table III), velocity-following plots (Fig. 7), torque-limit analysis during a jump (Fig. 8), policy-transition plots (Fig. 9), qualitative video comparisons against alternative RL formulations, and roughly ten hours of public deployment without a fall.","tokens_in":18262,"tokens_out":4779,"duration_ms":50462,"significance":"If the results hold, the paper provides a valuable demonstration that a character-driven, non-anthropomorphic biped can execute expressive, artist-directed motions while retaining robust dynamic mobility. The strengths include a clearly described divide-and-conquer policy architecture, detailed actuator system identification and domain randomization parameters, and a rare public-deployment record that substantiates the central claim. The paper is also notable for explicitly separating perpetual, periodic, and episodic motion types and for integrating show functions into the animation pipeline. The provided appendices give enough quantitative detail to make the approach reproducible in principle.","major_comments":[{"comment":"The mean absolute tracking errors in Table III are reported without any indication of the number of trials, the variance across runs, or the specific command and phase conditions under which they were measured. Since these numbers are the central quantitative support for the tracking claim, please report the number of episodes, the mean and standard deviation (or range), and the experimental protocol used to generate the data.","section":"Section VII-A, Table III"},{"comment":"The comparison with alternative RL formulations is entirely qualitative, relying on statements such as \"as shown in the video\" and \"visually identical motion.\" If the paper claims any comparative advantage, please provide quantitative metrics for the alternative policies (e.g., tracking error, foot clearance, joint acceleration, or energy consumption); otherwise, explicitly frame the comparison as qualitative so readers can calibrate the strength of the claim.","section":"Section VII-B"},{"comment":"The sim-to-real argument rests on actuator models identified on an isolated test bench and randomized only within isolated-actuator parameter ranges. The paper does not address in-situ effects such as shared-battery voltage sag, thermal drift, or coupled power draw during sustained high-power episodes, even though the presented jump reaches the actuator torque limits. Please add either measurements of battery voltage and motor temperature during representative shows or a discussion of the expected operating margin relative to the randomized envelope, so that the robustness claim is not overstated.","section":"Section V-E / Appendix B / Fig. 8"}],"minor_comments":[{"comment":"The statement that \"the robot is responsive and follows all commands closely\" would be stronger with a quantitative measure of velocity tracking error or latency, rather than only the visual comparison in the figure.","section":"Section VII-A, Fig. 7"},{"comment":"The hat notation for target state quantities (e.g., \\hat{p}, \\hat{q}) is used throughout Table I but is not explicitly defined in the text; please add a short explanation near Eq. (1).","section":"Section V-A / Table I"},{"comment":"The manuscript contains the watermark \"Disney Confidential - Do Not Distribute\" on every page; this is unusual for a journal submission and should be removed before final submission, or its presence must be explained in the cover letter.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong systems-integration contribution with compelling hardware evidence, and the central claim is defensible. However, the missing statistical detail in Table III and the unaddressed in-situ actuator boundary conditions are load-bearing for the robustness and tracking claims. The authors can likely address these points with additional analysis and a modest amount of new measurement, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a well-executed engineering integration paper, not a new methodology. The genuinely new pieces are the character-driven robot hardware and the tripartite policy decomposition (perpetual/periodic/episodic) paired with a real-time animation-blending puppeteering stack. If you work in entertainment robotics or sim-to-real character control, it's worth a careful read.\n\nThe hardware itself is new: a 15.4 kg non-humanoid biped with backward-bending knees, no ankle roll, passive rounded feet, and show functions (antennas, eyes, headlamp, speakers). The authors are explicit that the design is driven by creative intent, not function, and they back that with an iterative design loop between animation tools and rigid-body dynamics. The control side is a sensible divide-and-conquer: separate policies for standing, walking, and episodic motions, conditioned on low-dimensional commands, switched at runtime with phase-aware transitions. The policy-switch plots in Fig. 9 and the velocity tracking in Fig. 7 give concrete evidence that the system works.\n\nThe evaluation is honest about what it does and doesn't show. The MAE numbers in Tab. III have no error bars or trial counts, so they are suggestive rather than rigorous. The comparison to alternative RL formulations is qualitative (video only), and the 'believability' claim rests on anecdotal feedback from public deployments. None of that is fatal for an engineering demo, but it limits what we can conclude.\n\nOn the sim-to-real concern: it is real but not as sharp as the stress-test note suggests. The actuator models are identified on a test bench and don't explicitly model shared-battery voltage sag or thermal drift. But the paper reports about 10 hours of no-fall runtime across multiple robots and operators, which is strong aggregate evidence that the randomized policy transfers. The jump torque plot in Fig. 8 shows the policy operating right at the actuator limits, so the margin is likely thin under sustained high-power episodes. A careful referee should ask for voltage/thermal data or at least a discussion of why the deployed performance held. That is a boundary-condition question, not a refutation.\n\nWho this is for: researchers in legged locomotion, character animation, and HRI who want a concrete example of a full pipeline from artist-authored motion to reliable hardware. It deserves a serious referee. I'd ask for error bars, a quantitative baseline comparison, and ideally release of the animation processing code and actuator model parameters; without those, the reproducible core is thin. But the paper is coherent, self-aware, and the central claim—that this design and control workflow yields a robust, expressive character—is supported by the evidence presented.","headline":"A credible and well-integrated system paper for entertainment robotics; the evaluation is thin in places but the central engineering claim holds.","tokens_in":18857,"tokens_out":2437,"would_cite":false,"duration_ms":23412,"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 puppeteered bipedal robot can act out expressive, unscripted live shows, and the authors report roughly ten hours of public runtime with no falls.","keywords":["bipedal robot character","character-driven mechanical design","reinforcement learning locomotion","imitation learning from animation","sim-to-real transfer","robot puppeteering interface","animation engine","entertainment robotics"],"falsifier":"Run the physical robot through a fixed script of the same walking speeds, triggered animations, and walk-to-stand transitions used in the paper on a hard floor, and compare its falls, joint tracking errors, and foot-contact timing with the same script executed in simulation under identical commands. If the real tracking errors and fall rate exceed the simulated distribution substantially, or if the phase-synchronized policy switch becomes visibly discontinuous on hardware, the sim-to-real assumption is falsified.","tokens_in":17929,"feed_emoji":"🤖","tokens_out":7836,"duration_ms":78612,"temperature":0.7,"pith_summary":"This paper argues that expressive, artist-directed motion and dynamic balance can be unified in a single legged-robot system by co-designing the robot body with its motion repertoire, training separate reinforcement-learning policies to imitate artist-authored motions, and letting an animation engine blend those policies behind a two-joystick interface. A human operator can then puppeteer the robot through live, unscripted performances, switching between standing, walking, and one-shot emotional animations. The authors built a small bipedal character with creatively driven proportions and show functions such as actuated antennas and illuminated eyes, and deployed up to three robots in public settings, accumulating about ten hours of runtime without a single fall. If the workflow holds, custom expressive robot characters for entertainment and human engagement can be developed quickly without giving up dynamic mobility.","feed_headline":"Puppeteered bipedal robot keeps its feet through expressive live shows","feed_subtitle":"Artists drive the motion, the control stack keeps balance; public shows ran about ten hours with no falls.","key_machinery":"The key object is the path frame, a moving coordinate frame that anchors every artist-authored motion to the world and makes transitions between policies consistent. Each motion reference is stored in path coordinates and mapped to world coordinates by a generator; during standing the frame converges to the feet, during walking it integrates commanded velocities, and for episodic motions its trajectory is part of the artistic input. The path frame, together with a phase signal fed to the policy as a feature vector, lets separate policies be switched and blended seamlessly because every commanded pose is expressed in a common, robot-relative frame. The other load-bearing mechanism is the imitation-reward reinforcement-learning stack: a weighted sum of pose, velocity, joint, and contact rewards compares simulated motion to kinematic references, while actuator models and domain randomization bridge the gap between simulation and hardware.","core_discovery":"On the paper's own terms, the central discovery is that a single workflow—character-driven mechanical design, policies trained to imitate artist-authored motion references, and an animation engine that composes background, triggered, and joystick layers—can put expressive, dynamic performance on a physical bipedal robot and keep it believable in front of an audience. The design separates motion into three temporal types: perpetual standing, periodic walking driven by a phase signal and velocity commands, and episodic one-shot animations. Each type gets its own reinforcement-learning policy, and the animation engine's low-dimensional command signals, together with phase-aware switching, make the transitions look continuous to an outside observer. The evidence offered includes joint-tracking error tables, torque-limit checks during an episodic jump, responses to pushes and small obstacles, and roughly ten hours of public runtime across three robots with no falls.","pith_inferences":["The paper leaves implicit that the path-frame plus phase-signal command interface could serve as a standard API between animation tools and reinforcement-learning robot controllers, letting artists author for a robot without touching the control stack.","My extension would be a perceptual study measuring whether audiences rate a puppeteered character as more alive than the same motions played autonomously; the reported bystander questions about whether the robot can see suggest this is directly testable.","The ten-hour no-fall record is an operational result rather than a controlled benchmark, so a systematic stress test with scripted pushes and varied floor surfaces would separate policy robustness from operator skill."],"forward_implications":["A character robot can be developed in under a year from off-the-shelf actuators and 3D-printed structural parts, because the mechanical design is tuned to creative intent rather than extreme performance.","An operator without a robotics background can puppeteer expressive performances after training, because the interface separates gaze from posture and routes all commands through the animation engine.","Separate policies for standing, walking, and episodic motions can be swapped on the fly without visible discontinuities, as long as transitions are made phase-aware.","The same workflow can produce expressive characters with non-anthropomorphic morphologies, because the motion-reference format and path-frame interface are not tied to a human body plan.","Public deployments with up to three robots operating simultaneously accumulated about ten hours of runtime with zero falls."],"supporting_citations":[{"why":"Supplies the tracking-based imitation-reward approach that compares the simulated character to the reference pose.","marker":"[32]"},{"why":"Supplies the actuator model and system-identification procedure used to make simulation match real hardware.","marker":"[17]"},{"why":"Provides the sim-to-real actuator modeling and randomization methodology that underpins direct deployment without fine-tuning.","marker":"[42]"},{"why":"Supplies the GPU-accelerated physics simulation used to train the policies with domain randomization.","marker":"[26]"},{"why":"Provides the policy-gradient training algorithm used to optimize all control policies.","marker":"[38]"},{"why":"Supplies the animation-engine principles of composing and blending multiple animation sources based on external commands.","marker":"[44]"},{"why":"Supplies the procedural gait-generation tool that creates physically plausible periodic walking cycles for the character.","marker":"[14]"},{"why":"Supplies the state estimator that fuses IMU and actuator measurements on the deployed robot.","marker":"[11]"}],"fun_headline_variants":["Bipedal robot's character-driven design shines in live shows","RL control lets artist-driven bipedal robot perform live","Expressive bipedal robot: robust balance from RL policies","Character robot walks, jumps, and performs for ten hours"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that policies trained only in simulation transfer to the physical robot without fine-tuning; if the simulator's contact, actuator, or mass models are not faithful enough, the live performance could fail despite the reported ten hours.","fun_headline_variants_meta":{"raw":{"variants":["Bipedal robot's character-driven design shines in live shows","RL control lets artist-driven bipedal robot perform live","Expressive bipedal robot: robust balance from RL policies","Character robot walks, jumps, and performs for ten hours"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00016,"raw_usage":{"total_tokens":1185,"prompt_tokens":854,"completion_tokens":331,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":262}},"tokens_in":470,"tokens_out":331,"duration_ms":4342,"temperature":1.0,"reasoning_tokens":262,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:12:49.146770+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the physical robot through a fixed script of the same walking speeds, triggered animations, and walk-to-stand transitions used in the paper on a hard floor, and compare its falls, joint tracking errors, and foot-contact timing with the same script executed in simulation under identical commands. If the real tracking errors and fall rate exceed the simulated distribution substantially, or if the phase-synchronized policy switch becomes visibly discontinuous on hardware, the sim-to-real assumption is falsified.","supporting_citations":[{"cited_title":"Isaac gym: High performance gpu based physics simulation for robot learning","cited_arxiv_id":null,"evidence_quote":"Supplies the GPU-accelerated physics simulation used to train the policies with domain randomization."},{"cited_title":"van Breemen","cited_arxiv_id":null,"evidence_quote":"Supplies the animation-engine principles of composing and blending multiple animation sources based on external commands."},{"cited_title":"Hopkins, Georg Wiedebach, Kyle Cesare, Jared Bishop, Espen Knoop, and Moritz B ¨acher","cited_arxiv_id":null,"evidence_quote":"Supplies the procedural gait-generation tool that creates physically plausible periodic walking cycles for the character."},{"cited_title":"Contact-aided invariant extended kalman filtering for robot state estimation","cited_arxiv_id":null,"evidence_quote":"Supplies the state estimator that fuses IMU and actuator measurements on the deployed robot."}],"review_version":1}