PaperlayerView exact revision on arXiv

Source-linked AI summary

BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion

Qiayuan Liao, Takara E. Truong, Xiaoyu Huang, Yuman Gao, Guy Tevet, Koushil Sreenath, C. Karen Liu

arXiv:2508.08241v4cs.RO

TL;DR

Humanoid robots still face difficulty achieving natural, human-like movement, while prior approaches make local stylistic adjustments without restoring overall coordination. BeyondMimic presents a scalable pipeline for learning agile skills from human motion and extends imitation toward broader motion synthesis, with direct hardware transfer and demonstrated skill composition. Its guidance can destabilize denoising during mode switching or in high-variance states.

  • Problem

    Natural, human-like movement remains difficult, and local stylistic adjustments may not restore overall human-like coordination.

  • Method

    BeyondMimic is a scalable pipeline that learns directly from human motion to reproduce agile behaviors and uses score-function-based guidance in its diffusion process.

  • Results

    BeyondMimic learns hundreds of skills, transfers the same policy directly to hardware, and moves beyond imitation toward broader motion synthesis.

  • Takeaways & Limitations

    The framework demonstrates scalable human-like motor-skill acquisition and motion synthesis beyond imitation.

  • Takeaways & Limitations

    Guidance can destabilize the denoising process during mode switching or in high-variance states.

Abstract

from arXiv · show

The human-like form of humanoid robots positions them uniquely to achieve the agility and versatility in motor skills that humans possess. Learning from human demonstrations offers a scalable approach to acquiring these capabilities. However, prior works either produce unnatural motions or rely on motion-specific tuning to achieve satisfactory naturalness. Furthermore, these methods are often motion- or goal-specific, lacking the versatility to compose diverse skills, especially when solving unseen tasks. We present BeyondMimic, a framework that scales to diverse motions and carries the versatility to compose them seamlessly in tackling unseen downstream tasks. At heart, a compact motion-tracking formulation enables mastering a wide range of radically agile behaviors, including aerial cartwheels, spin-kicks, flip-kicks, and sprinting, with a single setup and shared hyperparameters, all while achieving state-of-the-art human-like performance. Moving beyond the mere imitation of existing motions, we propose a unified latent diffusion model that empowers versatile goal specification, seamless task switching, and dynamic composition of these agile behaviors. Leveraging classifier guidance, a diffusion-specific technique for test-time optimization toward novel objectives, our model extends its capability to solve downstream tasks never encountered during training, including motion inpainting, joystick teleoperation, and obstacle avoidance, and transfers these skills zero-shot to real hardware. This work opens new frontiers for humanoid robots by pushing the limits of scalable human-like motor skill acquisition from human motion and advancing seamless motion synthesis that achieves generalization and versatility beyond training setups.

Introduction

Humanoid control must address high-dimensional balance and coordination challenges while achieving agility, naturalness, and versatility. BeyondMimic combines scalable motion tracking with diffusion-based online optimization to learn diverse human motions and compose them for unseen tasks.

  • Introduction: High-dimensional joint coordination and floating-base balance make natural humanoid movement difficult.These challenges arise from the many joints and actuators and the difficulty of stable balancing.
  • Introduction: Model-based simplifications can produce unnatural motions and restrict robots to a small subset of their full motion range.Examples include constant center-of-mass height and persistently bent knees.
  • Introduction: Reinforcement-learning approaches often require hand-crafted, behavior-specific rewards that are costly to redesign and may produce unnatural movements.Naturalness and human-likeness are difficult to express as explicit optimization terms.
  • Introduction: Prior motion-tracking policies achieve agile and natural skills but rely on motion-specific tuning and do not generalize beyond motions seen during training.Adversarial motion-prior policies can generalize beyond specific clips but are generally not reusable across tasks and require retraining from scratch.
  • Introduction: Hierarchical and generative approaches improve reuse but can sacrifice agility and naturalness, suffer planner–controller mismatches, or generalize poorly to implicit objectives.Obstacle avoidance and long-horizon navigation are cited as examples of difficult implicit objectives.
  • Introduction: BeyondMimic uses a compact, motion-generalizing tracking formulation and a unified diffusion controller to compose skills and solve unseen tasks without task-specific training.The framework is presented as combining scalable human-motion learning with diffusion-based online optimization and enabling zero-shot transfer to real humanoids.

Diverse Skills

BeyondMimic scales motion tracking across diverse human motions using one formulation and shared hyperparameters, transferring a broad repertoire of behaviors from simulation to hardware.

  • Diverse Skills: Approximately 2.5 hours of diverse human motions supported training a broad set of skills.The passage states that these motions were successfully validated in high-fidelity simulation.
  • Diverse Skills: Thirty representative clips totaling 15 minutes transferred reliably from simulation to the physical robot.The deployed behaviors included static, balance-critical, dynamic, stylized, and expressive motions.
  • Diverse Skills: The framework retained agility and stylistic details across diverse motions without motion-specific tuning.Challenging clips were trained jointly with multiple other skills, including references exceeding three minutes.

Human-level Agility

BeyondMimic demonstrates human-level agility in dynamic and contact-rich behaviors, including acrobatics, ground interactions, and repeated expressive movements under challenging conditions.

  • Human-level Agility: Human-level agility was demonstrated in highly dynamic motions across outdoor environments with soft soil, leaves, and uneven ground.The evaluation included environmental conditions that differed from standard flat, rigid surfaces.
  • Human-level Agility: The robot completed complex acrobatic and martial-arts-inspired motions under deformable and unstable contacts absent during training.The demonstrated sequences included aerial cartwheels.
  • Human-level Agility: 31 m/s2 peak acceleration and up to 20 rad/s pelvic angular velocity were measured during an aerial cartwheel.The reported mean pelvic angular velocity was 7.01 rad/s, and landings required minimal recovery.
  • Human-level Agility: The robot performed two consecutive cartwheels, crawling, ground-up jumps, and repeated celebration jump-turns without loss of stability or style.It repeated the celebration five times in succession, whereas prior work reported one execution.

Natural, Human-like Behaviors

BeyondMimic produces human-like behaviors and supports inference-time skill composition through diffusion guidance. Its motions show perceptual and biomechanical similarities while enabling unseen-task control without retraining.

  • Natural, Human-like Behaviors: The robot reproduced human-like walking and running with comparable ground-reaction-force shapes and aligned loading and push-off timing.Walking showed double peaks, while running showed single peaks; walking force peaks were sharper than human profiles.
  • Natural, Human-like Behaviors: The lack of a toe joint limited rollover and push-off during stance, making the difference more pronounced in walking than running.The running limitation was less pronounced because contact occurred once per step and for a shorter duration.
  • Natural, Human-like Behaviors: The robot remained compliant during external disturbance, stabilized when held, and smoothly resumed walking after release.It avoided stiff or exaggerated reactions.
  • Natural, Human-like Behaviors: Diffusion guidance enables velocity and waypoint commands, keyframe inpainting, and flexible task composition without retraining.The model iteratively refines predicted trajectories toward specified objectives using a learned prior over feasible motor skills.
  • Natural, Human-like Behaviors: A learned prior over diverse feasible motor skills lets simple task-specific costs trigger behaviors for unseen tasks without numerous regularization terms.This distinguishes the approach from online trajectory optimization as described in the passage.

Command-conditioned Locomotion

BeyondMimic supports command-conditioned locomotion through velocity and waypoint inputs, producing multimodal, smooth, and perturbation-robust human-like behaviors. It also maintains stable control over extended distances.

  • Command interfaces: The policy performs zero-shot commanded locomotion from desired velocities or waypoint targets.Waypoint navigation reaches goals from varying initial positions, while joystick control tracks commanded directions omnidirectionally.
  • Robustness: The controller continues pursuing objectives under large external perturbations, including kicks.This demonstrates robustness while maintaining commanded locomotion behavior.
  • Robustness: Over 50 m of continuous running indicates stable control over extended distances.Average velocity tracking error was 12.14% in walking and 13.65% in running.
  • Multimodal locomotion: The policy produces distinct human-like gaits under similar low-velocity commands, including walking and light jogging.It also transitions smoothly from walking to running using only desired velocity input.
  • Multimodal locomotion: Rare and unlabeled transitions in the motion data emerge naturally when required by the task.The controller inferred smooth skill transitions through command context rather than explicit skill labels.

Motion Inpainting and Task Transitioning

Motion inpainting enables smooth insertion and composition of agile skills, while the diffusion controller transitions between motion and task specifications. These capabilities support coherent long-horizon behaviors and task switching.

  • Motion Inpainting: Sparse future keyframes guide smooth intermediate motions, enabling online insertion, transition, and composition of agile skills.The setup begins from joystick-controlled walking and injects desired cartwheel keyframes during execution.
  • Motion Inpainting: The diffusion policy inpaints cartwheel keyframes into a temporally coherent continuous trajectory.After completing the cartwheel, the controller smoothly returns to command-conditioned walking.
  • Agile skill transitions: The model transitions from command-conditioned locomotion into contact-rich and highly dynamic motions, including standing recovery, spin-kicks, and flip-kicks.These behaviors are learned from human motion tracking and include emergent transitions.
  • Task Transitioning: Online optimization enables task transitions across different specifications, including velocity following and motion inpainting.The framework demonstrated switching between these tasks and long-horizon sequences involving walking, running, and cartwheels.

Task Composition

BeyondMimic composes learned humanoid skills and task objectives without retraining, combining human-like motion tracking with online planning and control. It also transfers diverse behaviors directly to hardware while preserving agility and naturalness.

  • Task Composition: Task composition addresses the combinatorial growth of goal combinations and tasks that are difficult to cover fully during training.Multiple objectives can instead be evaluated and optimized at inference time without enumerating every possibility.
  • Task Composition: Simple task-specific costs can be flexibly summed, enabling compositions of unseen tasks without retraining.The framework demonstrated scene-aware navigation and obstacle avoidance through composed objectives.
  • Obstacle Avoidance: Signed-distance obstacle costs provide trajectory gradients during denoising, steering predictions away from obstacles.The robot detoured around an obstacle and reached the target waypoint; joystick tracking could replace waypoint tracking while mitigating collisions under moderate off-goal inputs.
  • Scalable Motion Tracking: The framework learns hundreds of diverse motions with one principled RL formulation and shared hyperparameters, without extensive heuristics.Moderate domain randomization and careful system implementation were sufficient instead of extensive torque, delay, and contact-force heuristics.
  • Hardware Transfer: The same policy transfers directly to hardware without motion- or robot-specific tuning while retaining agility and natural appearance.The learned repertoire includes human-like, stylized, and expressive behaviors, and the work reports substantial advancement over published humanoid-control results.
  • Unified Control: A unified model synthesizes and composes atomic behaviors into coherent novel trajectories and solves downstream tasks.Classifier guidance adjusts future trajectories at inference time, coupling planning and control within one model for navigation, joystick commands, and disturbance recovery.
  • Unified Control: The model preserves smooth human-like behavior while navigating obstacles, following joystick-driven velocity commands, and recovering under external disturbances.Its task-agnostic, unlabeled-data training supports adding motion data without redefining objectives or retraining for each task.
  • Broader Implications: BeyondMimic shifts humanoid control from manually tuned motion-specific policies toward predictive, generalizable behavior models.The proposed direction targets generalization across motions and environments while unifying human-like behavior with planning and control.

Limtations and Future Works

BeyondMimic scales human-motion tracking and guided diffusion across diverse behaviors and unseen tasks, while exposing limitations in prediction, guidance stability, state estimation, and fine-grained control.

  • Limitations: The diffusion model can become trapped in repetitive motion patterns, while larger guidance weights may destabilize denoising during mode switching or high-variance states.
  • Limitations: Generated trajectories inherit state-estimation errors, although latent diffusion provides some robustness to noisy observations.
  • Limitations: The system predicts a 0.64-second horizon, supporting reactive control and local obstacle avoidance but not long-horizon planning toward distant goals or obstacles.
  • Limitations: Guided diffusion is stable after a gait orbit is established but tends to stumble at motion starts and ends.
  • Limitations: Guidance-based optimization works well for coarse-grained objectives but less well for fine-grained ones and still requires lightweight guidance-weight tuning.
  • Results: The framework learns diverse human motions with shared hyperparameters, scales to hundreds of skills without manual tuning, and transfers both stages to physical robots.
  • Method: A unified diffusion model supports expressive skill composition and online task optimization, synthesizing diverse motions for unseen tasks.

Tracking Objective

The tracking objective reproduces a retargeted human reference on a humanoid while using relative body poses to preserve motion style and tolerate global drift.

  • Tracking Objective: The method tracks a human-retargeted reference motion for deployment on a real humanoid robot.
  • Tracking Objective: Reference generalized positions and velocities are converted through forward kinematics into poses for each robot body link.
  • Tracking Objective: A compact target-body set is selected for tracking, including the robot’s end-effectors, to avoid redundancy among closely spaced links.
  • Tracking Objective: Relative body poses are tracked instead of global poses to preserve motion style while allowing global drift from perturbations and sim-to-real mismatch.
  • Tracking Objective: An anchor body, typically the root or torso, expresses desired poses of other bodies in an anchor-centered frame.

𝑏, 𝑇anchor

The tracking controller uses relative anchor-centered errors, lightweight motion-agnostic regularization, proprioceptive observations, and adaptive training procedures to improve robustness and scalability.

  • Tracking Objective: The objective uses yaw-aligned, height-preserving transforms while keeping desired twists unchanged, preserving motion style under benign global drift.
  • Reward Design: Tracking errors cover position, orientation, linear velocity, and angular velocity across selected target bodies.
  • Reward Design: The reward combines a unified task-space tracking term with three broadly applicable penalties for joint limits, action smoothness, and self-collision.
  • Observation and Action: The robot-centric observation includes motion phase, anchor pose error, IMU twist, joint states, and the previous action, which helps suppress high-frequency jitter.
  • Observation and Action: Normalized joint-position setpoints shape desired torques through low-level PD control rather than serving as precise position plans.
  • Observation and Action: High-impedance control can amplify sensor noise, reduce passive impact compliance, and obscure implicit torque information on hardware.
  • Training Robustness: Compact domain randomization improves scalable training, while adaptive sampling prioritizes difficult trajectory segments before gradually restoring uniform coverage.

Versatile Humanoid Control via Guided Diffusion

BeyondMimic models state–action trajectories in a latent diffusion space and applies inference-time guidance to compose human-like skills for unseen objectives without retraining.

  • Predictive Control: A latent state–action diffusion model predicts short future trajectories so task objectives can steer future states and generate corresponding actions.
  • Predictive Control: Jointly modeling states and actions with causal consistency bridges state-space task objectives and action-space policy outputs without relying on forward dynamics models.
  • Latent Diffusion Model: The model uses a VAE for smooth motion representations and a state–latent diffusion model trained on unlabeled, task-agnostic trajectories.
  • Latent Diffusion Model: At inference, an unseen cost steers the diffusion model toward desired trajectories, whose latent outputs are decoded into joint-level actions.
  • Latent Diffusion Model: Latent diffusion addresses irregular action distributions and inference latency by providing smooth motion representations with a lightweight decoder.
  • Training: The two-stage learning process first encodes motion intents with a conditional VAE, then trains an LDM on state–latent trajectories collected from VAE rollouts.
  • Online Optimization: Classifier guidance enables training-free inference-time optimization for joystick control, obstacle avoidance, and motion inpainting.

Validation of the method

The validation studies examine motion-tracking design choices, sim-to-real deployment sensitivity, and the framework’s ability to execute diverse and unseen behaviors. Results support scalable tracking, agile hardware demonstrations, natural locomotion, and guided task composition.

  • Motion-tracking ablations: Rot6D produced markedly better sim-to-real performance than quaternion and axis–angle rotation representations.Quaternion baselines had higher tracking errors, while the axis–angle baseline failed once.
  • Deployment sensitivity: 2 ms of latency increased velocity error, 5 ms caused one failure, and 10 ms caused failures in two of three trials.The authors identify carefully engineered real-time deployment as important for robust sim-to-real transfer.
  • Motion-tracking ablations: Adaptive sampling was critical for difficult motion segments, with three of four motions failing without it after 30k iterations.For easier motions, it also halved training iterations from 4k to 2k.
  • Hardware validation: Thirty distinct motion clips were deployed on hardware, spanning static, highly dynamic, athletic, and stylized behaviors.Demonstrated behaviors included aerial cartwheels, spin kicks, and flip kicks, with tracking transferred reliably to hardware.
  • Natural locomotion: The robot exhibited comparable weight-normalized ground-reaction-force shapes, contact forces, and timing to humans during walking and running.A user study found the proposed policies significantly more preferred than Unitree’s native controller for human-like naturalness.

Supplementary materials

The supplementary materials specify actuator-inertia modeling, motion-tracking rewards, training procedures, domain randomization, and controller-gain choices. They also report tracking comparisons across natural frequencies and describe task-cost formulations for downstream control.

  • Actuator modeling: Reflected rotor inertia is modeled as a physical joint contribution proportional to the square of the gear ratio, rather than as a solver-stability parameter.The supplementary discussion states that reflected inertia can alter apparent inertia and dynamic response.
  • Actuator modeling: Dual-actuated ankle and waist roll/pitch joints are modeled with joint armature equal to twice the actuator value under an effectively 1:1 linkage.Most other joints are single-actuated, so their joint armature equals the actuator armature.
  • Actuator modeling: Reflected inertia contributes a large, often dominant, fraction of effective axis inertia, especially toward distal joints, constraining controller bandwidth and gain selection.The stated practical consequence is that actuator-reflected inertia matters more than link-subtree inertia for these joints.
  • Motion tracking: Motion tracking combines position, orientation, linear-velocity, and angular-velocity errors with smoothness and other regularization terms.Tracking errors are averaged over target bodies and normalized with Gaussian-shaped exponential functions.
  • Training procedure: Training uses asymmetric actor–critics, adaptive motion-phase sampling based on smoothed bin failure rates, domain randomization, and periodic velocity perturbations.The reference trajectory is divided into one-second bins, with exponential moving-average failure rates guiding reset sampling.
  • Controller gains: A moderate natural frequency produces the best global tracking performance and stronger local tracking than 5 Hz and ASAP, while 25 Hz has slightly lower local pose error but noticeably higher stiffness.The supplementary results also state that the heuristic gain formula surpasses hand-tuned gains used in prior work.
Loading 2508.08241v4…