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

ARCHITECTURE

Figure S1: Motion Tracking Overview. Our actor-critic architecture for motion tracking. The actor network takes observations: motion phase 𝝍, anchor pose error 𝐞anchore_{anchor}, IMU twist 𝒱imuV_{imu}, relative joint positions 𝜽−𝜽0{

THE PROBLEM

Before BeyondMimic, humanoid robotCore ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. controlControl & PlanningControlThe method used to make the robot move the way you want. faced two stuck problems. First, motion tracking—the process of making a robotCore ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. physically follow human motion capture data—was either brittle or required constant hand-tweaking per motion type. Systems like physics-informed neural networks and reinforcement learningImitation & Reinforcement LearningReinforcement Learning (RL)Teaching a robot through trial and error using rewards. approaches would either produce jerky, unnatural movements or need task-specific rewardImitation & Reinforcement LearningRewardA score that tells the robot how well it is doing. shaping for each new behavior. Second, motion imitation was siloed: you'd train one policyCore ConceptsPolicyThe rule or model that maps observations or states to actions. for walking, another for jumping, another for manipulationManipulation & TasksManipulationUsing a robot arm or hand to move or interact with objects.. Composing these into new behaviors or adapting them to novel goals at test time was nearly impossible. Prior work (like mocap-guided RLImitation & Reinforcement LearningReinforcement Learning (RL)Teaching a robot through trial and error using rewards. or motion-specific controllers) could handle individual skills but couldn't scale across diverse, high-frequency dynamic motions, and certainly couldn't generalize to tasks outside their training distributionData, Distributions & Training IssuesTraining distributionThe kinds of examples the model saw during training.. The industry standard was: one motion type = one trainingRobot LearningTrainingThe process of fitting a model using data or experience. run = one hyperparameter set. BeyondMimic breaks this.

HOW IT WORKS

1

Compact Motion-Tracking MDP Formulation

The authors designed a motion tracking setup that removes motion-specific tuning by treating kinematic tracking as a standard RLImitation & Reinforcement LearningReinforcement Learning (RL)Teaching a robot through trial and error using rewards. problem, but with an elegant twist: they use a single, shared reward functionImitation & Reinforcement LearningReward functionThe rule that defines how rewards are assigned. across all 14 different motions (cartwheels, sprints, kicks, etc.). The key innovation is a compact formulation that tracks reference jointMovement, Mechanics & Robot BodyJointA movable connection between robot parts. angles while letting the policyCore ConceptsPolicyThe rule or model that maps observations or states to actions. discover balance, timing, and ground contactMovement, Mechanics & Robot BodyContactPhysical interaction between the robot and an object or surface. implicitly. This is elegant because it sidesteps the usual approach of hand-crafting cost functions for each skillModern Robot LearningSkillA reusable behavior like grasp, push, place, or open drawer. type. The same MDP setup and hyperparameters work for both a slow cartwheel and a fast sprint—proof that the formulation captures something fundamental about how humanoids move. The motion tracking pipeline outputs trajectories that are both dynamically feasible (the robotCore ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. can actually execute them) and human-like (they look natural, not robotic).

Show more ↓

Figure 1: Overview of the proposed versatile humanoid control framework. (A) Scalable and robust learning from human motions with agile, human-like behaviors via motion tracking. (B) Versatile control over unseen downstream tasks with diverse learned motor skills via guided diffusion.

Figure 2: Diverse motion tracking policies deployed on a real humanoid robot. We demonstrate accurate and robust tracking across a wide spectrum of motions, ranging from static to highly dynamic and from athletic feats to stylized motions, showing the scalability of our framework. In total, 30 disti

Figure 5: Command-conditioned locomotion via guided diffusion. (A) Visualization of diffusion process under joystick control. Starting from Gaussian noise, the distribution progressively converges to optimize for a right-turn command. (B) Waypoint navigation. From multiple start points, the robot re

Figure 7: Overview of the framework. (A) Scalable learning of diverse human motions via motion tracking. (i) The target motion is re-anchored to the current pose to allow drift and recovery, which is critical for successful sim-to-real transfer. (ii) Diverse motions are tracked with RL using a singl

prep IMG 2643

prep IMG 2669

prep IMG 2661

prep IMG 4081

2

Distillation into a Unified Latent Diffusion Model

Once the team had mastered tracking 14 diverse motions, they faced the composition problem: how do you blend these into a single policyCore ConceptsPolicyThe rule or model that maps observations or states to actions. that can switch between them or combine them on the fly? They distilled all these motion-tracking policies into a single latent diffusion model—essentially a generative model that learns the underlying structure of human motion. Diffusion models are powerful here because they naturally capture uncertainty and multimodality: there are many valid ways to reach a waypoint or avoid an obstacle. By encoding policies in a learned latent spaceRobot LearningLatent spaceA compressed internal representation space inside a model., the model becomes amenable to test-time guidance. This is where the real versatility unlocks: you can now specify goals at test time using simple cost functions (like 'reach this waypoint' or 'avoid this obstacle'), and the diffusion model generates a trajectoryCore ConceptsTrajectoryA sequence of states or actions over time. that solves the taskCore ConceptsTaskThe job the robot is supposed to complete, such as pick-and-place, navigation, or drawer opening. while staying faithful to learned human motion priors. You're not retraining; you're reusing.

Show more ↓

Figure 3: Human-level Agility exhibited in a highly dynamic acrobatic motion. (A) Real-world forest demonstration showing one continuous motion including: 180∘180^{} side flip (aerial cartwheel), two consecutive 180∘180^{} side spin kicks, and a 360∘360^{} flip kick. (B) Robot’s orien

Figure 4: Human-like naturalness in walking and running motion. (A) Natural running on a sports field. (B) Natural walking on a sports field. (C) Comparison of weight-normalized GRF profiles for human vs. humanoid during walking and running, showing comparable shapes, contact forces, and timing. (D)

Figure S2: Extra Ablation On PD Gains Ablating on PD Gains, We find that following our heuristics formula with a moderate natural frequency yields the best global tracking performance and surpasses the hand-tuned gains used in prior work (?). Although increasing the natural frequency can further red

prep C1921

prep C1909

prep C1876

prep C1915

3

Classifier Guidance for Zero-Shot Task Generalization

The final trick is classifier guidance—a diffusion-specific technique for steering a generative model toward specific objectives at test time. The authors train a lightweight classifier or cost function that scores how well a candidate motion achieves a goalCore ConceptsGoalThe desired outcome or target state for a robot task. (e.g., reaching a waypoint, avoiding an obstacle, responding to joystick commands). During inferenceRobot LearningInferenceUsing a trained model to make predictions or choose actions., they nudge the diffusion sampling process toward high-scoring trajectories. This is powerful because it lets the model solve tasks it never saw during trainingRobot LearningTrainingThe process of fitting a model using data or experience. without any fine-tuningModern Robot LearningFine-tuningTaking a pretrained model and adapting it to a specific robot or task.. Joystick teleoperationImitation & Reinforcement LearningTeleoperation (teleop)A human remotely controlling the robot, often to collect demonstrations.? Never trained on that explicitly. Obstacle avoidanceNavigation & LocomotionObstacle avoidanceMoving while avoiding collisions with obstacles. with specific geometries? Works zero-shotModern Robot LearningZero-shotDoing a new task without task-specific training.. Motion inpainting (fill in missing frames)? Works too. The classifier guidance acts as a bridge between the motion priors learned from human data and the new taskCore ConceptsTaskThe job the robot is supposed to complete, such as pick-and-place, navigation, or drawer opening. at hand. Critically, this all transfers to real hardware without retraining—the real robotCore ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. runs the same diffusion model and classifier guidance that worked in simulationSimulation & Sim-to-RealSimulationA virtual environment where robots can be trained or tested..

Show more ↓

teaser

MORE DEMONSTRATIONS

FIGURES

Figure 6: Task transition and composition. (A) Motion inpainting with future keyframes. (i) Desired future keyframes with 0.20.2 s intervals. (ii) Starting from walking, the robot smoothly completed the cartwheel by inpainting both the transition and the intermediate motion between the keyframes. (i

KEY RESULTS

PERFORMANCE COMPARISON

WHY DEVELOPERS SHOULD CARE

LIMITATIONS

BeyondMimic's main limitations are practical rather than conceptual. The motion tracking pipeline still relies on external mocap or state estimationPerception & SensingState estimationCombining noisy sensor data to estimate the robot’s true state. to function well—the real robotCore ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. demos used mocap for waypoint and obstacle location, which isn't available in many real-world scenarios. The diffusion model, while versatile, still operates on a motion primitiveModern Robot LearningPrimitive / action primitiveA simple reusable low-level movement or control building block. foundation: it can't invent entirely new classes of movement not present in the trainingRobot LearningTrainingThe process of fitting a model using data or experience. data, and performance likely degrades gracefully as you request behaviors farther from the learned manifold. The paper doesn't deeply analyze failure modes—when does classifier guidance fail? How sensitive is performance to cost function design? Also, the approach assumes access to high-quality human mocap data for the relevant skillModern Robot LearningSkillA reusable behavior like grasp, push, place, or open drawer. space, which limits its applicability to domains where good mocap isn't available (underwater robots, aerial drones, etc.). Finally, computational cost of diffusion sampling at 10+ Hz for real-time controlSimulation & Sim-to-RealReal-time controlProducing actions fast enough for live robot control. isn't discussed, which matters for deploymentSimulation & Sim-to-RealDeploymentPutting the trained system on a real robot..

WHAT COMES NEXT

The natural next steps are (1) removing the mocap dependency by developing better proprioceptive and visual state estimationPerception & SensingState estimationCombining noisy sensor data to estimate the robot’s true state. so the robotCore ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. can operate fully autonomously in real environments, (2) extending the framework to full-body manipulationManipulation & TasksManipulationUsing a robot arm or hand to move or interact with objects. by including arm and hand tracking in the mocap imitation pipeline, (3) scaling to longer-horizon reasoning (the current work is good for 1-5 second behaviors; what about multi-minute tasks with planningControl & PlanningPlanningFiguring out what the robot should do before or during movement.?), and (4) exploring online adaptation—can the diffusion model update its motion priors from real-world experience without retraining from scratch? There's also an interesting direction in hierarchical diffusion: use one diffusion model to select high-level motion primitives and another to refine them for specific tasks. Finally, if sim-to-realSimulation & Sim-to-RealSim-to-real (sim2real)Transferring a policy trained in simulation to a real robot. worked this well, there's probably a paper in understanding *why*—what properties of the learned representations made them transfer? That understanding could unlock even more robust frameworks for other robotCore ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. morphologies.

RELATED PAPERS