Multi-Terrain Humanoid Locomotion Challenge
ICCV 2025 Workshop on Human-Robot-Scene Interaction and Collaboration
This challenge focuses on developing robust locomotion controllers for humanoid robots to navigate complex, unstructured terrains - a critical capability for real-world deployment in homes, disaster zones, and natural environments.
Using the high-performance Isaac Gym simulator, participants will create locomotion policies that enable humanoids to traverse diverse surfaces including slopes, stairs and rubble terrain. The competition provides standardized robots and terrains, allowing you to focus on advanced locomotion strategies.
Your goal is twofold:
The final score combines both distance coverage and efficiency metrics.
All participant teams need to register by emailing challenge-terrain@outlook.com with team information. You will recieve an unique token for evaluating your solution.
Subject: [ICCV2025-HRSIC]Humanoid Locomotion Challenge Registration
Body:
Here is our registration information of [Team Name]:
1. Point of Contact: [Team Member 1], [Affiliation], [Email]
2. [Team Member 2], [Affiliation]
...
Whether to open-source the code (for awards): [Yes/No]
Additional comments: [Any additional information or questions]
If you have any questions, feel free to reach out! We may also send updates regarding the challenge to registered teams.
We innovatively utilize a client-server architecture for policy evaluation. The server hosts the simulated evaluation environments, while clients evaluate their policies via our provided REST API.
The REST API is designed to have the same interface as gym, e.g., env.reset() and env.step(action). Thus no modifications are needed for existing gym-compatible policies. The results will be shown on the public leaderboard.
Export and submit a JIT-compiled model. Use our exporter template as a guide, you may implement your own JIT exporter and produce a serialized JIT model compatible with the evaluation runner.
Thanks to our sponsors at Unitree and Fourier: đ„ First Prize ($1000) đ„ Second Prize ($500) đ„ Third Prize ($300)
Your policy must process standardized observations and output specific control signals.
The policy receives the following observations at each timestep, in the exact order as concatenated in the observation buffer:
| Observation | Dimension |
|---|---|
| Base angular velocity | 3 |
| Base orientation | 4 (quaternion) |
| Current yaw error | 1 |
| Next yaw error | 1 |
| Forward velocity command | 1 |
| Joint positions | Nj |
| Joint velocities | Nj |
| Previous actions | Nj |
| Terrain height samples | 132 |
Where Nj is equal to the specified robotâs number of degrees of freedom. Detailed observation specifications will be provided in our challenge repository. Stay tuned!
Your policy must output:
There are 3 robots available for the challenge:
Policies will be evaluated on three specialized scenarios (detailed in Scenarios):
We compute two metrics for each evaluation episode:
Episode Score = Completion Rate Ă 0.9 + (1 - Efficiency Score) Ă 0.1
The score will be averaged across all scenario episodes for each robot. The award is selected based on the total score summed over all robots.
Our open-source Terrain Benchmark is accessible in this repository.
We recommend using the competition_only branch, which allows training, playing, and evaluating models directly without any code modification.
In contrast, the master branch provides additional observations for research flexibility, but models trained there cannot be evaluated directly on the challenge website.
We have designed different scenarios to evaluate the different performance of robots. All terrain used for evaluation comes from the single terrain model of the terrain module provided in the code.
