Source-linked AI summary
Humanoid Parkour Learning
Ziwen Zhuang, Shenzhe Yao, Hang Zhao
TL;DR
Humanoid parkour demands varied skills, active perception, and autonomous decisions across challenging terrain without fixed motion references. The paper trains a unified whole-body-control policy with fractal-noise terrain and two-stage visual distillation, achieving multi-skill parkour and onboard deployment. Real-world experiments show autonomous skill selection, robustness to arm-action overrides, and successful operation across indoor and outdoor settings.
Problem
Humanoid parkour requires diverse skills and egocentric perception, while prior learning methods depend on constrained rewards, motion references, or substantial terrain-specific data.
Method
The framework trains a unified oracle parkour policy with fractal-noise terrain and then uses DAgger to distill onboard depth vision while preserving turning commands.
Results
The policy autonomously selects appropriate skills for challenging parkour tasks using onboard depth vision and performs across 10 simulated skills, outperforming feet-air-time reward training.
Takeaways & Limitations
The framework supports humanoid parkour without motion references and remains robust when arm actions are overridden for potential mobile-manipulation use.
Takeaways & Limitations
The method relies on manually constructed terrain and is difficult to generalize to unseen terrains without further training.
Abstract
from arXiv · showhide
Parkour is a grand challenge for legged locomotion, even for quadruped robots, requiring active perception and various maneuvers to overcome multiple challenging obstacles. Existing methods for humanoid locomotion either optimize a trajectory for a single parkour track or train a reinforcement learning policy only to walk with a significant amount of motion references. In this work, we propose a framework for learning an end-to-end vision-based whole-body-control parkour policy for humanoid robots that overcomes multiple parkour skills without any motion prior. Using the parkour policy, the humanoid robot can jump on a 0.42m platform, leap over hurdles, 0.8m gaps, and much more. It can also run at 1.8m/s in the wild and walk robustly on different terrains. We test our policy in indoor and outdoor environments to demonstrate that it can autonomously select parkour skills while following the rotation command of the joystick. We override the arm actions and show that this framework can easily transfer to humanoid mobile manipulation tasks. Videos can be found at https://humanoid4parkour.github.io
1 Introduction
Humanoid parkour requires a unified system that perceives terrain, selects skills, and executes varied maneuvers without relying on fixed obstacle orders or extensive motion references. This paper introduces a vision-based learning framework for multiple human-capable terrains using onboard computation and zero-shot sim-to-real transfer.
- Parkour requires active perception, memory, autonomous skill selection, and powerful whole-body execution across obstacles such as platforms, gaps, and hurdles.
- Existing humanoid parkour demonstrations rely on extensive offline engineering and may be limited to a single obstacle order.
- Diverse skills are difficult to unify because reward designs encouraging foot raising or preventing foot drag can restrict leg motion.
- Human or animal motion references lack egocentric perception and require large datasets spanning terrains and embodiments, limiting scalability.
- The proposed system covers at least 10 human-capable terrain types without motion references, using fractal noise, two-stage training, and DAgger-based vision distillation.
2 Related Work
Legged locomotion research has progressed from model-based control toward reinforcement learning, but limited exteroception has constrained many learned systems to low-speed locomotion.
- Model-based legged locomotion methods require substantial mechanical design and dynamics modeling, generally targeting robust walking on planar surfaces.
- Reinforcement learning became applicable to legged locomotion with general physical simulators, but many learning algorithms remain limited to low-speed locomotion because they lack exteroception.
3 Training Method and Robot System
The training system builds a unified oracle parkour policy from planar walking, diverse procedurally varied terrains, and targeted safety rewards, then distills visual perception for onboard deployment. It preserves turning commands while replacing terrain scans with depth-image perception and accelerating DAgger distillation across GPUs.
- The pipeline trains a zero-shot end-to-end policy using onboard depth vision and proprioception, organized into three stages beginning with a plane-walking policy.
- Fractal noise and multiple reward components train walking and foot raising, while virtual-obstacle penetration and footstep guidance add redundancy and stepping accuracy.
- The oracle policy uses scandot terrain embeddings, a GRU state estimator, an MLP actor, proprioception, and PPO to predict target joint actions.
- Training Forward Parkour from Planar Walking: The turning command is computed from the goal-heading error, with forward motion disabled when the angular error reaches at least π/2.
- Train an Oracle Parkour Policy: 10 terrain types are arranged with controllable difficulty, and an auto-curriculum moves robots toward harder or easier sub-terrains according to task completion.
- Distill Visual Perception from Scandots: A student CNN encodes 48 × 64 depth images, while simulated depth clipping, Gaussian noise, artifacts, and real-world filtering bridge the simulation-to-reality gap.
- Multi-Processes Distillation Acceleration: DAgger distillation uses four GPU processes, with one trainer and three trajectory collectors labeling student data through the oracle policy.
4 Experiment Results
The experiments evaluate the parkour system in simulation and real-world indoor and outdoor settings, including terrain robustness, onboard-vision skill selection, and arm-action override. Results cover fractal-noise training, quantitative obstacle performance, balance during arm motion, and multi-GPU distillation.
- Real-World Quantitative Results: Our parkour policy achieves the best performance in four difficult real-world tasks against blind walking and Unitree H1’s built-in MPC controller.Each testing configuration uses 10 trails, with controller foot-height adjustments applied to help overcome some obstacles.
- Fractal Noise with No Motion Prior is Effective: Fractal-noise training outperforms the “feet airtime” reward across all 10 simulated skills in success rate and average moving distance.Success rate is computed at maximum terrain difficulty, and average moving distance measures travel before falling or stopping.
- Onboard Vision is Crucial: Onboard depth vision lets the robot select the proper skill for challenging parkour tasks while following a directional command.The real-world quantitative experiment filters out operator failure because the policy follows a directional command.
- Distillation and Acceleration: The student policy distilled from scratch struggles with balance, while one-GPU distillation reaches only half the performance of fully accelerated distillation after 24 hours.The comparison measures transitions experienced by the student policy and success rate after 24 hours of training.
- Observations in Locomotion Behavior: During jumping, the policy swings its right arm to keep balance, and teleoperated arm motion still permits balance during parkour execution.Arm override is explicitly described as out-of-distribution for the parkour policy.
5 Conclusion, Limitations and Future Directions
The paper presents a motion-reference-free humanoid parkour system and demonstrates robustness to arm-action override for potential manipulation use. It identifies unseen-terrain generalization and more difficult manipulation as directions requiring further training and tuning.
- Conclusion: The system applies to humanoid parkour without complex motion references and supports robust arm-action override for manipulation tasks.The paper reports indoor parkour and outdoor locomotion tests, with videos provided online.
- Limitations: The method relies on manually constructed terrain and is difficult to generalize to unseen terrains without further training.This is the paper’s stated limitation concerning terrain scope.
- Future Directions: More difficult manipulation skills require further tuning to avoid disrupting the visual system.Future work will investigate additional training that coordinates manipulation while preserving extreme locomotion ability.
C Reinforcement Learning Training Details
The training details specify reward design, PPO configuration, terrain visualization, and obstacle attributes used for reinforcement learning and evaluation.
- Reward Terms: Torque penalties are weighted by each joint’s kp factor to account for differing motor capacities and kp values.The paper states that none of these reward terms use motion references for humanoid locomotion.
- Optimization: PPO trains the policy, while the value network uses the same architecture with ground-truth simulated linear velocity.The value network differs from the policy network in its velocity input.
- Terrain and Obstacles: Scandots visualize terrain information, including wave and discrete terrain, and Table 6 lists critical attributes for each parkour obstacle.Fractal noise is added to terrain height fields so robots learn to raise their feet while moving around.
E Policy Architecture
The policy architecture combines recurrent state estimation and action generation with terrain encoding, using different terrain encoders for oracle and deployable student policies. Deployment processes depth vision and proprioception at separate rates on Unitree H1 hardware.
- Policy Architecture: The oracle policy uses recurrent state estimation, a recurrent actor, and an MLP terrain encoder, while the student replaces terrain encoding with a CNN depth encoder.Both policies retain the same RNN-MLP state estimator and RNN-MLP actor architecture.
- Configuration: The architecture and deployment configuration are specified through policy structure, obstacle ranges, and joint kp, kd, and torque tables.These tables define the policy structure, critical obstacle properties, and joint control factors.
- Deployment Pipeline: Depth images are resized from 480 × 640 to 48 × 64, with the visual encoder running at 10Hz and the rest of the network at 50Hz.Proprioception arrives through Cyclone DDS at 500Hz, and the encoder sends image latents to the parkour network.
- Hardware Setup: The Unitree H1 deployment uses an Intel RealSense D435i, an onboard Intel Core i7 12-core NUC, 19 joints, ROS2, and Cyclone DDS.The depth encoder sends embeddings to the recurrent network at 10Hz.
G Further Results Compared with feet air time reward
The simulation comparison evaluates success rate and average moving distance across connected terrain tracks of increasing difficulty.
- Each terrain is formed from three connected sub-tracks with linearly increasing difficulty.
- Each sub-track is 4.8m long, giving each terrain a total length of 14.4m.
- The comparison reports success rate and average moving distance in simulation.
H Results Compared with Different Distillation Methods
The distillation experiments compare methods using the same oracle policy and fixed 24-hour simulation runs, showing disadvantages for single-GPU and randomly initialized student training.
- 24-hour distillation runs compare three methods using the same oracle policy as teacher.
- Single-GPU distillation produces worse success rate and average moving distance than the comparison method.
- Randomly initialized student-policy distillation also produces worse performance.
A Experiment Videos
The supplementary materials document the real-world evaluation, training setup, policy architecture, terrain design, and deployment implementation for the humanoid parkour system.
- Experiment Videos: The system is evaluated through indoor parkour tests and outdoor locomotion tests.
- Training Details: Training uses uniformly sampled domain-randomization attributes across 4096 robots during reinforcement learning.
- Reinforcement Learning Training Details: The reward design includes projected gravity and torque penalties weighted by each joint’s kp factor, without motion-reference terms.
- Terrain and Parkour Setup: The training setup implements 10 terrain types and uses scandots to perceive terrain information.
- Policy Structure: The oracle policy uses an RNN-MLP state estimator, RNN-MLP actor, and MLP terrain encoder, while the student replaces the terrain encoder with a CNN depth encoder.
- Real-World Deployment: On the Unitree H1, depth vision runs at 10Hz, the parkour network at 50Hz, and motor torque control at 1000Hz.
H Results Compared with Different Distillation Methods
The distillation study examines why multi-GPU acceleration matters for humanoid parkour policy training and finds that limited training throughput or random initialization harms learning.
- The study investigates why multi-GPU acceleration is necessary for distilling a humanoid parkour policy.
- Three distillation settings use the same oracle policy as teacher and are evaluated with 20 simulation trials per setting.
- After 24 hours, a randomly initialized student policy still struggles to walk steadily.
- A one-GPU run trains on 4.147 × 10^6 transitions versus 432 × 10^6 transitions for the 4-GPU variant in 24 hours.
- The one-GPU variant detects different obstacle types, but its legs do not raise high enough to overcome them.