Source-linked AI summary
Hiking in the Wild: A Scalable Perceptive Parkour Framework for Humanoids
Shaoting Zhu, Ziwen Zhuang, Mengjie Zhao, Kun-Ying Lee, Hang Zhao
TL;DR
Humanoid hiking needs proactive perception, but existing mapping-based and scalable-control approaches face estimation, motion, and training challenges. Hiking in the Wild uses a single-stage end-to-end policy with depth and proprioception, safety-aware foothold learning, and feasible target sampling. Field experiments demonstrate traversal of challenging terrain at speeds up to 2.5 m/s, while the authors identify limited camera coverage and multitask mode collapse as remaining boundaries.
Problem
Humanoid hiking requires proactive terrain perception because reactive proprioception can miss hazards, while existing exteroceptive and end-to-end approaches face estimation or scalability challenges.
Method
The framework maps raw depth and proprioception directly to joint actions and combines Terrain Edge Detection with Foot Volume Points and Flat Patch Sampling.
Results
2.5 m/s maximum running speed was achieved during zero-shot sim-to-real deployment, alongside successful traversal of high platforms and discrete gaps.
Takeaways & Limitations
The framework demonstrates robust humanoid hiking across stairs, slopes, uneven grass, and gaps using a scalable perceptive control approach.
Takeaways & Limitations
The single forward-facing depth camera limits backward and lateral movement, while simultaneous training across terrains and gait modes can cause mode collapse and degraded performance.
Abstract
from arXiv · showhide
Achieving robust humanoid hiking in complex, unstructured environments requires transitioning from reactive proprioception to proactive perception. However, integrating exteroception remains a significant challenge: mapping-based methods suffer from state estimation drift; for instance, LiDAR-based methods do not handle torso jitter well. Existing end-to-end approaches often struggle with scalability and training complexity; specifically, some previous works using virtual obstacles are implemented case-by-case. In this work, we present \textit{Hiking in the Wild}, a scalable, end-to-end parkour perceptive framework designed for robust humanoid hiking. To ensure safety and training stability, we introduce two key mechanisms: a foothold safety mechanism combining scalable \textit{Terrain Edge Detection} with \textit{Foot Volume Points} to prevent catastrophic slippage on edges, and a \textit{Flat Patch Sampling} strategy that mitigates reward hacking by generating feasible navigation targets. Our approach utilizes a single-stage reinforcement learning scheme, mapping raw depth inputs and proprioception directly to joint actions, without relying on external state estimation. Extensive field experiments on a full-size humanoid demonstrate that our policy enables robust traversal of complex terrains at speeds up to 2.5 m/s. The training and deployment code is open-sourced to facilitate reproducible research and deployment on real robots with minimal hardware modifications.
I. INTRODUCTION
Hiking in unstructured environments requires humanoids to perceive terrain proactively rather than rely only on reactive proprioception. Hiking in the Wild addresses this with scalable end-to-end perception, foothold safety, feasible target sampling, and direct real-robot deployment.
- Motivation: Blind locomotion reacts robustly to contact forces but can miss deep gaps or high steps before collisions cause catastrophic falls.The paper frames proactive perception as necessary for safer hiking in unknown terrain.
- Related challenges: Mapping-based exteroception depends on precise state estimation, while LiDAR suffers from low frequency and motion distortion during dynamic movement.These limitations motivate alternatives that avoid explicit localization and map reconstruction.
- Approach: The framework maps raw depth and proprioception directly to joint actions through a single-stage reinforcement-learning policy with high-frequency depth processing.The design targets scalable perceptive locomotion without external state estimation.
- Safety mechanism: Terrain Edge Detection with Foot Volume Points penalizes unsafe edge penetration so the policy learns foothold placement on safer, flatter surfaces.This mechanism addresses humanoid sensitivity to partial edge contacts without explicit trajectory planning.
- Command generation: Flat Patch Sampling selects reachable flat terrain regions as navigation targets to reduce reward hacking from arbitrary velocity commands.The strategy is intended to encourage meaningful traversal rather than spinning in place.
- Evaluation: Field experiments show robust traversal of stairs, slopes, uneven grass, and discrete gaps, including high-speed locomotion.The reported evaluation covers previously unseen environments and multiple challenging terrain types.
B. Perceptive Foothold Control
The paper formulates perceptive humanoid locomotion as a POMDP and supplies the policy with proprioceptive, command, action-history, and depth observations. Depth processing includes temporal history, while the policy outputs target joint positions executed through PD-controlled torques.
- Problem formulation: Perceptive humanoid locomotion is formulated as a partially observable Markov decision process optimized with PPO.The formulation defines observation, action, termination, and reward components for policy training.
- Observation space: The actor observation combines base angular velocity, projected gravity, velocity commands, joint positions, joint velocities, the previous action, and depth images.The observation includes both proprioceptive and perceptive signals.
- Observation space: A historical sequence of depth images provides temporal context for the actor, while the critic additionally receives noise-free observations and base linear velocity.The asymmetric actor-critic architecture supports simulation training.
- Action execution: The policy outputs target joint positions, which are converted into actuator torques through PD control.The computed torques are applied to execute the desired motion.
- Depth processing: Depth processing distinguishes synthetic and real-world inputs through the Fsim and Freal pipelines shown in the processing figure.The figure contrasts the top-row synthetic-data processing with bottom-row real-world-data processing.
3) Termination Criteria:
The framework defines termination constraints that exclude physically unfeasible states and accelerates training convergence, while its depth pipeline synthesizes sensor-realistic observations from simulated geometry.
- Termination criteria: Episodes terminate on time-out, terrain-boundary violations, illegal torso contact, unstable orientation, or insufficient root height.These constraints prevent exploration of physically unfeasible states and accelerate training convergence.
- Reward structure: The reward combines task, regularization, safety, and AMP-style components for command tracking, energy efficiency, constraint satisfaction, and natural locomotion.
- Depth synthesis: The ray-caster computes each pixel’s radial distance from the first intersection between its ray and terrain or robot geometry.
- Depth synthesis: Orthogonal depth is obtained by projecting radial distance onto the camera’s principal axis using the ray direction and forward-facing unit vector.
- Sensor modeling: Simulation depth processing sequentially applies cropping and resizing, range-dependent Gaussian noise, disparity-artifact masking, blur, clipping, normalization, and OOD dropout.
- Sensor modeling: The real-sensor pipeline applies cropping and resizing, depth inpainting for zero-valued regions, and Gaussian blur.
3) Temporal Depth Aggregation via Strided Sampling:
The policy aggregates depth observations with a strided temporal history, extending temporal context while processing fewer frames and supporting anticipation during fast maneuvers.
- Temporal sampling: A history buffer stores m depth frames sampled with temporal stride ℓ, with a single-frame training delay modeling sensor latency.
- Temporal sampling: The historical sequence is H_t = {I_t−k·ℓ | k = 0, 1, . . . , m−1}.
- Temporal sampling: The policy perceives a horizon of (m−1)·ℓ steps while processing only m discrete frames.
- Purpose: Sparse extended history captures terrain-profile trends and relative velocity without redundant overlap from consecutive high-frequency frames.
- Purpose: This strided history supports anticipating obstacles and adjusting gait during high-speed maneuvers.
C. Terrain Edge Contact Penalization
Terrain Edge Detection and Foot Volume Points create a geometric safety penalty that discourages unstable, edge-adjacent foot contacts and promotes centered footholds.
- Terrain Edge Detection: The edge detector identifies sharp terrain edges by thresholding dihedral angles between adjacent faces.
- Terrain Edge Detection: Raw edges are filtered for noise and short segments are greedily concatenated to reduce the number of geometric primitives.
- Foot Volume Points: Volume points are distributed within each foot’s collision manifold and queried against a spatial collision grid for penetration depth.
- Penalization: The penalty uses each volume point’s penetration offset and foot velocity to discourage unstable interactions near terrain edges.
- Penalization: The formulation penalizes high-velocity impacts or scraping motions more heavily, encouraging stable footholds.
D. Position-based Velocity Command
Uniformly sampled velocity commands can produce reward hacking in which the robot turns in circles instead of crossing obstacles; prior remedies modify rewards or use goal commands.
- Reward hacking: Uniform random velocity commands can reward turning in circles rather than traversing obstacles.
- Existing remedies: Previous approaches address this behavior by modifying velocity-tracking rewards or using goal-based commands.
- Existing remedies: Reward tuning with randomly sampled commands makes reaching maximum performance difficult, while pure goal commands often lack control over robot speed.
1) Target Generation via Flat Patch:
Flat Patch Sampling identifies reachable, sufficiently level terrain regions as navigation targets, replacing arbitrary targets with feasible destinations. The method samples locations, ray-casts local heights, and accepts patches whose height variation stays below a threshold.
- Target validity: Flat patches are accepted when terrain within radius r has maximum height difference below threshold δ.The procedure uses ray-casting to evaluate surrounding terrain levels.
- Sampling procedure: The algorithm samples random 2D positions and ray-casts heights within radius r around each candidate.Accepted candidates are added to the patch set with their average height.
- Terrain coverage: The resulting flat patches provide examples of stable terrain targets across different terrains.
2) Position-based Velocity Command Generation:
Position-based commands derive forward and turning velocities from reachable target locations, while separate motion-prior training supplies walking and running style. History-based discriminator transitions and stabilized AMP rewards support temporal motion learning.
- Velocity generation: Velocity commands are generated from the target’s relative position, with forward and angular gains controlling linear and rotational motion.Lateral velocity is set to zero because the forward camera supports forward locomotion and heading alignment.
- Turning commands: A subset of flat-terrain agents receives vx = 0 and random ωz commands to learn in-place turning.Combining turning commands with position-based commands improves maneuverability across environments.
- Motion priors: AMP uses synthetic, human, and running motion data to provide stable and agile reference behaviors.Walking and running policies are trained separately using datasets totaling 379.62 s and 1.54 s, respectively.
- Temporal modeling: The discriminator receives a history of n frames rather than a single state pair to capture temporal motion features.
- AMP stabilization: MSE loss with quadratic rewards provides smoother, non-saturating gradients than binary cross-entropy and log-based rewards.The stated purpose is to avoid vanishing gradients and support more stable convergence.
IV. EXPERIMENTS
Experiments evaluate zero-shot deployment, edge-aware foothold safety, and component contributions across challenging terrains and walking or running gaits. Real-world tests show high-speed traversal, high success across scenarios, and sustained walking without falls.
- Evaluation scope: The evaluation covers stairs, high platforms, grassy ramps, and discrete gaps using both walking and running gaits.The experiments address training and deployment, edge-aware safety, and the contributions of perception, AMP, and command generation.
- Real-world traversal: 2.5 m/s maximum running speed was achieved while traversing platforms up to 32 cm and discrete gaps 50 cm wide.These results were obtained through direct zero-shot sim-to-real deployment on the humanoid robot.
- Perception: 60 Hz depth perception provided high-frequency environmental feedback for rapid posture and terrain adjustments during dynamic tasks.
- Reliability: High success rates were maintained across nearly all tested terrain and gait scenarios despite real-world sensory noise and physical discrepancies.Each terrain-gait combination was evaluated in 10 trials.
- Long-duration stability: 4 minutes of continuous walking across staircases and flat surfaces were completed without falls or human intervention.
C. Edge-aware Penalization Mechanism
The edge-aware mechanism encourages footholds away from terrain boundaries by penalizing unsafe foot-volume penetration. It improves simulated success and landing area, reduces deployment risks, and generalizes edge detection to new terrain types without manual tuning.
- Safety mechanism: Terrain Edge Detection and Volumetric Point Penetration Penalization encourage feet to maintain safety margins from terrain edges.The mechanism is designed to preserve stable contact during foothold placement.
- Simulation results: The method outperforms No Edge in both success rate and landing area.Landing area measures the portion of the foot manifold contacting the terrain surface.
- Edge-dense terrain: The observed landing-area gap indicates substantial improvement specifically in edge-dense areas.Mean landing-area values are generally high because episodes include considerable flat-ground time.
- Deployment safety: Stepping on terrain edges can cause slippage or unpredictable contact dynamics that drive the robot into OOD states and immediate falls.The proposed mechanism reduces these risks by incentivizing larger landing areas and edge avoidance.
- Scalability: The edge detector identifies edges on Stones and Stakes without manual feature engineering or parameter tuning.This is reported as zero-shot performance across new terrain types.
D. Ablation Study on Training Recipe
The ablation study compares the full training recipe with variants removing temporal depth history, pose-based commands, MoE, or AMP, evaluating success and reaching time across terrains. The full method achieves the highest success rate, supporting the combined use of the proposed components.
- Ablation settings: The ablation compares the full framework with variants using single-frame depth, uniform commands, a vanilla MLP, or no motion prior.These variants remove depth history, pose-based command generation, MoE, or AMP, respectively.
- Evaluation protocol: Each policy is tested for 10,000 time-steps with 1,000 simulated robots across terrain types.Evaluation uses Success Rate and Mean Reaching Time.
- Results: The full method achieves the highest success rate across the ablation configurations.The result is reported as evidence that integrating all proposed components is essential for traversing complex geometries.
- Overall recipe: The framework combines volumetric edge-aware penalization with position-based velocity command generation for high-dynamic behavior on complex terrains.The paper reports experiments in simulation and real-world deployment, including agile running and stable walking.
- Scope of the recipe: Training diverse terrains and gait modes simultaneously can cause mode collapse and lower performance than specialized policies.The authors identify advanced multitask reinforcement learning as a possible direction for improving a unified policy.