PaperlayerView exact revision on arXiv

Source-linked AI summary

ZEST: Zero-shot Embodied Skill Transfer for Athletic Robot Control

Jean Pierre Sleiman, He Li, Alphonsus Adu-Bredu, Robin Deits, Arun Kumar, Kevin Bergamin, Mohak Bhardwaj, Scott Biddlestone, Nicola Burger, Matthew A. Estrada, Francesco Iacobelli, Twan Koolen, Alexander Lambert, Erica Lin, M. Eva Mungai, Zach Nobles, Shane Rozen-Levy, Yuyao Shi, Jiashun Wang, Jakob Welner, Fangzhou Yu, Mike Zhang, Alfred Rizzi, Jessica Hodgins, Sylvain Bertrand, Yeuhi Abe, Scott Kuindersma, Farbod Farshidian

arXiv:2602.00401v1cs.ROcs.AI

TL;DR

Robust, human-like whole-body control remains difficult for agile, contact-rich robot behaviors because existing approaches often require extensive per-skill engineering and complex control pipelines. ZEST trains single-stage reinforcement-learning policies from heterogeneous motion references and deploys them zero-shot across robots, demonstrating broad hardware capabilities while retaining important scope limitations.

  • Problem

    Robust, human-like whole-body control for agile, contact-rich behaviors remains challenging because sim-to-real discrepancies and partial observability complicate deployment.

  • Method

    ZEST trains a single-stage motion-imitation policy from MoCap, ViCap, and animation while using minimal observations, adaptive sampling, and a model-based assistive wrench.

  • Results

    ZEST demonstrates zero-shot hardware skills across Atlas, Unitree G1, and Spot using heterogeneous motion references, including dynamic multi-contact and acrobatic behaviors.

  • Takeaways & Limitations

    The demonstrated framework provides a unified interface for transferring varied biological-motion references into whole-body control across multiple robot embodiments.

  • Takeaways & Limitations

    The study does not test multi-skill generalization to unseen motions and assumes proprioception with flat, non-slippery terrain.

Abstract

from arXiv · show

Achieving robust, human-like whole-body control on humanoid robots for agile, contact-rich behaviors remains a central challenge, demanding heavy per-skill engineering and a brittle process of tuning controllers. We introduce ZEST (Zero-shot Embodied Skill Transfer), a streamlined motion-imitation framework that trains policies via reinforcement learning from diverse sources -- high-fidelity motion capture, noisy monocular video, and non-physics-constrained animation -- and deploys them to hardware zero-shot. ZEST generalizes across behaviors and platforms while avoiding contact labels, reference or observation windows, state estimators, and extensive reward shaping. Its training pipeline combines adaptive sampling, which focuses training on difficult motion segments, and an automatic curriculum using a model-based assistive wrench, together enabling dynamic, long-horizon maneuvers. We further provide a procedure for selecting joint-level gains from approximate analytical armature values for closed-chain actuators, along with a refined model of actuators. Trained entirely in simulation with moderate domain randomization, ZEST demonstrates remarkable generality. On Boston Dynamics' Atlas humanoid, ZEST learns dynamic, multi-contact skills (e.g., army crawl, breakdancing) from motion capture. It transfers expressive dance and scene-interaction skills, such as box-climbing, directly from videos to Atlas and the Unitree G1. Furthermore, it extends across morphologies to the Spot quadruped, enabling acrobatics, such as a continuous backflip, through animation. Together, these results demonstrate robust zero-shot deployment across heterogeneous data sources and embodiments, establishing ZEST as a scalable interface between biological movements and their robotic counterparts.

INTRODUCTION

ZEST addresses the challenge of robust, general whole-body control by training a unified reinforcement-learning policy from heterogeneous motion data and deploying it across robots without per-skill engineering. Its streamlined recipe supports zero-shot transfer while reducing reliance on contact labels, extensive reward shaping, history windows, state estimators, and multi-stage pipelines.

  • Humanoid robots need broad, coordinated human-like movement to operate effectively in environments designed around the human body.
  • Existing approaches face sim-to-real difficulty in contact-rich, near-limit motions and often add state estimation, history conditioning, or handcrafted pipeline components.
  • ZEST trains a single-stage policy from MoCap, ViCap, and keyframe-animation references while avoiding contact labels, extensive reward shaping, future-reference windows, observation histories, and state estimators.
  • ZEST applies across Atlas, Unitree G1, and Spot, and evaluates dynamic multi-contact skills from motion capture, video-derived behaviors, and animation.
  • The policy uses reference and proprioceptive observations, outputs residual joint targets, and relies on adaptive sampling plus a model-based assistive wrench for difficult long-horizon motions.

RESULTS

Hardware experiments show ZEST tracking diverse motion references across Atlas, Unitree G1, and Spot, including locomotion, multi-contact behaviors, dance, box interaction, and other dynamic skills. The results also expose higher tracking discrepancies for contact-rich or dynamically aggressive motions, while video-derived skills remain expressive despite noisy references.

  • ZEST evaluates specialized policies on Atlas, Unitree G1, and Spot using MoCap, ViCap, and keyframe-animation references, with hardware trials reported across embodiments.
  • MoCap-Derived Human-Like Skills on Hardware: MoCap references produce walking, jogging, rolling, crawling, cartwheeling, breakdance, table-tennis, and other behaviors on Atlas and G1.
  • MoCap-Derived Human-Like Skills on Hardware: The same formulation handles multi-contact motions without prescribed contact schedules or upright base constraints, including rolls, crawls, cartwheels, and breakdance.
  • MoCap-Derived Human-Like Skills on Hardware: Foot-only gaits show lower MAE(q) and MAD(R), whereas dynamic multi-contact skills show higher orientation and angular-velocity discrepancies.
  • ViCap-Derived Human-Like Skills on Hardware: ViCap experiments maintain fluid, temporally coherent motion over long sequences despite handheld-camera reconstruction noise and visible pose jitter and foot skidding.
  • ViCap-Derived Human-Like Skills on Hardware: G1 box skills and ballet require whole-body contacts, sustained balance, precise foot placement, and operation near the robot’s limits.

Keyframe-Animated Skills for Humanoids and Quadrupeds on Hardware

ZEST uses keyframe animation to extend skills beyond natural human movement and demonstrates transfer across humanoid and quadruped platforms. Its adaptive training strategies improve coverage of difficult motion segments, while comparisons show advantages over MPC for dynamic behaviors and arbitrary-body contacts.

  • Keyframe-Animated Skills for Humanoids: Keyframe animation enables Atlas to exploit robot-specific continuous joints and achieve configurations unavailable to human motion capture, including a handstand invert.The handstand uses the robot’s back-yaw joint while maintaining stable hand support.
  • Keyframe-Animated Skills for Quadrupeds: The same framework enables Spot to perform foreleg handstand balance, continuous backflips, barrel rolls, and a playful happy-dog motion despite the larger morphology gap.These skills are trained from animation references.
  • Simulation-Based Evaluation and Ablation Studies: Adaptive sampling increases training on failure-prone trajectory segments, while assistive-wrench scaling supports difficult long-horizon motions such as cartwheel-to-backflip.Failure rates raise sampling probabilities, and the hardest segment receives more training than the easiest animated squat.
  • Simulation-Based Evaluation and Ablation Studies: Under full domain randomization, evaluations use 10,000 rollouts per reference and report success after 10 and 20 hours, with ablations testing curriculum, sampling, windows, actions, and actor–critic observations.The ablation setup also evaluates robustness to pushes, observation noise, masses, friction, and initial-state perturbations.
  • Simulation-Based Evaluation and Ablation Studies: Longer observation or future-reference windows hinder convergence under the same budget, while removing critic privileged observations or using absolute actions degrades performance.The longer-window result may depend on the reported architecture and hyperparameters.
  • Simulation-Based Benchmarking Against MPC: Compared with MPC, the learned policy performs better on dynamic behaviors such as cartwheel and jog and accommodates contacts involving knees, torso, and forearms without explicit contact labels.Several MPC behaviors fail because of contact annotation, saturation, or difficulty regulating extended contacts.

DISCUSSION

ZEST connects heterogeneous motion references to zero-shot whole-body control across multiple robot embodiments. The authors identify generalization to unseen motions, non-flat environments, and automated system identification as open directions.

  • ZEST trains reinforcement-learning policies from heterogeneous references and deploys them zero-shot, avoiding contact labels, state estimators, extensive reward shaping, and multi-stage training.
  • The framework demonstrates dynamic multi-contact behaviors on Atlas, video-based skill transfer on Unitree G1, and policy transfer across distinct morphologies including Spot.
  • The study does not test multi-skill policy generalization to unseen motions that are close to, but not identical with, the training distribution.
  • The current proprioceptive formulation assumes flat, non-slippery terrain, leaving explicit perception of uneven or compliant environments for future work.
  • Sim-to-real transfer depends on reasonable actuator modeling and armature-dependent gains, while fully automated system identification remains open.

MATERIALS AND METHODS

This section introduces the technical components of ZEST, centered on a goal-conditioned MDP and its path from reference-motion data to hardware deployment.

  • The technical presentation covers ZEST’s architecture, goal-conditioned MDP, robot and reference-motion modeling, dataset curation, training setup, and hardware deployment.

Systems and Modeling

ZEST uses progressively simplified parallel-linkage actuator models to make closed-chain robot dynamics practical for simulation and control. The system combines approximate armature modeling with a more detailed actuator model for highly dynamic Spot behavior.

  • The evaluation spans Atlas, Unitree G1, and Spot, while parallel-linkage actuators create computationally expensive closed-chain dynamics in humanoid simulation.
  • Parallel-Linkage Actuator Models: The locally projected model treats support links as massless while retaining motor-armature and main-chain inertia, producing configuration-dependent effective armature.
  • Parallel-Linkage Actuator Models: The dynamic armature model uses a Jacobi approximation for coupled joints, approximating off-diagonal effects with only a minor, transient error.
  • Parallel-Linkage Actuator Models: The nominal armature model fixes values at one configuration, avoiding an approximately 20% timestep-update slowdown and supporting PD-gain design.
  • Actuator Refinement: Spot’s more detailed actuator model incorporates power limiting, motor magnet saturation, transmission inefficiency, and friction to address triple-backflip sim-to-real requirements.
  • Modeling Progression: The modeling progression runs from an exact closed-loop model through locally projected and Jacobi-based approximations to fixed-configuration armatures.

Reference Motion Pipeline

ZEST builds a reference-motion library from MoCap, monocular video, and keyframe animation, then maps human-centric motion to each robot through kinematic retargeting. The pipeline uses spacetime optimization to align motion while preserving physically consistent timing and scale.

  • The reference library combines MoCap, ViCap, and keyframe animation without contact labels, relying on kinematic information and retargeting to the target robot skeleton.
  • MoCap supplies high-fidelity natural motion, while ViCap reconstructs 3D human motion from a single camera using camera-scene estimation and human-pose estimation.
  • Spacetime optimization retargets MoCap and ViCap by aligning root motion and corresponding bones, penalizing joint velocities, and jointly optimizing scale and time resampling.

Training Setup and MDP Formulation

ZEST formulates motion imitation as a goal-conditioned MDP in which a deployment-compatible policy tracks reference targets using onboard proprioception and minimal history. Training combines generic rewards, privileged critic information, adaptive sampling, assistive-wrench curriculum, and moderate domain randomization to support robust dynamic behavior.

  • Training Setup and MDP Formulation: ZEST models reference tracking as a goal-conditioned MDP, where goals may be target snippets, latent embeddings, or phase and trajectory identifiers.The policy maximizes discounted return over randomized rollouts and goal distributions.
  • Actor and Critic Observations: The actor uses proprioception and target references, while the critic additionally receives privileged dynamics, contact, end-effector, and curriculum signals unavailable during deployment.This asymmetric design improves value estimation while keeping the policy compatible with hardware observations.
  • Actions and Joint-Level Control: The policy outputs residual joint targets added to reference positions before joint-level PD control, with per-DoF scales allocating more assistance to demanding actuators.PD gains use approximate armature-based second-order models and balance responsiveness against deployment robustness.
  • Reward Terms: Generic tracking, regularization, and survival rewards support diverse behaviors without task-specific terms or reference contact labels.Regularization penalizes abrupt actions, large accelerations, and joint-limit violations while survival rewards longer episodes.
  • Early Terminations: Early termination ends rollouts after excessive contact forces or large reference deviations, steering exploration away from highly undesirable trajectories.These conditions are used to avoid wasting samples on failed behavior.
  • Domain Randomization: Domain randomization injects observation noise, random pushes, friction variation, and link-mass variation to reduce overfitting to the nominal simulator.The randomization is intentionally moderate because excessive variability can degrade performance.
  • Adaptive Reference State Initialization: Reference State Initialization samples trajectory phases at reset, exposing the policy to relevant states and decoupling episode length from demonstration length.Uniform sampling can oversample mastered regions, motivating adaptive bin sampling based on recent tracking failures with a floor probability for every bin.
  • Assistive Wrench Automatic Curriculum: An automatic assistive-wrench curriculum gradually increases difficulty for highly dynamic motions that otherwise terminate immediately or converge slowly under RSI alone.The curriculum targets behaviors with large base rotations and angular rates, including handstands, cartwheels, and backflips.

Hardware Deployment Pipeline

ZEST uses an automated, framework- and hardware-agnostic deployment pipeline to support safe and consistent policy releases across multiple robot platforms.

  • Hardware Deployment Pipeline: The deployment workflow exports training checkpoints in ONNX format, records them in a registry, and attaches detailed metadata for release management.The described pipeline is designed for frequent physical-robot testing across heterogeneous hardware.

SUPPLEMENTARY MATERIALS

The supplementary materials document nomenclature, implementation details, modeling procedures, dataset construction, curriculum methods, and additional evaluation videos.

  • Supplementary Sections: Supplementary sections cover actuator and Spot modeling, motion datasets, adaptive RSI sampling, and the assistive-wrench curriculum.These sections provide implementation details beyond the main training description.
  • Supplementary Videos: Supplementary movies show retargeted references, multi-skill and MPC evaluations, box-climbing tests, and the effect of PLA modeling on sim-to-real transfer.The listed videos include both simulation benchmarks and hardware repeatability and robustness tests.

Supplementary Material

The supplementary material develops simulator-compatible models for closed-chain actuators and training mechanisms that focus assistance and sampling on difficult motion segments.

  • Projected Model: The exact Projected Model represents closed-chain actuator dynamics through main and support chains connected by kinematic mappings and interaction forces.The main chain contains the primary robot kinematic tree, while the support chain contains the parallel-link support structure.
  • Projected Model: The exact model is incompatible with standard rigid-body simulator structures, motivating approximations tailored for efficient reinforcement-learning training.The derivation also assumes unactuated parent joints and no external forces in the simplified analysis.
  • Dynamic Armature Model: The Dynamic Armature Model approximates coupled closed-chain dynamics with diagonal armature terms and a fictitious torque representing off-diagonal effects.The approximation uses previous-timestep joint acceleration and introduces only transient error when the mass matrix is diagonally dominant.
  • Nominal Armature Model: The Nominal Armature Model fixes armature values at one configuration, reducing simulation overhead while also providing a basis for fixed low-level PD gains.This permits projected motor armatures to be added to the main kinematic tree when training begins.
  • Model Validation: The Locally Projected, Dynamic Armature, and Nominal Armature models closely match the exact model, whereas the Simplest Model has larger but qualitatively similar motion errors.The Simplest Model excludes fictitious torque from the Nominal Armature model.
  • Assistive-Wrench Curriculum: The assistive-wrench curriculum couples assistance to adaptive sampling: difficult bins receive stronger initial assistance, which decays as similarity improves while sampling frequency remains elevated until failures subside.A floor probability preserves coverage, and assistance reaches zero at the desired similarity threshold.
Loading 2602.00401v1…