kaus4004/robust-humanoid-ppo-results

RobustHumanoid PPO Results

This dataset contains RobustHumanoid benchmark results for a Stable-Baselines3 PPO policy evaluated on Humanoid-v4 under out-of-distribution MuJoCo physics.

The benchmark evaluates a fixed policy over a 20 x 20 grid of ground friction and actuator torque scale values, with 5 episodes per grid point.

Benchmark Configuration

• Benchmark: RobustHumanoid
• Environment: Humanoid-v4
• Policy name: ppo_humanoid
• Policy type: Stable-Baselines3 PPO
• Episodes per setting: 5
• Seed: 0
• Deterministic evaluation: true
• Friction grid: 0.1, 0.2, ..., 2.0
• Torque scale grid: 0.1, 0.2, ..., 2.0
• Total physics settings: 400

Results

Metric	Value
Robustness area	1.02429634557675
Mean return	323.139785583621
Median return	307.4571266094464
Nominal return	315.47489843056195
Worst-case return	261.0482641669373
Best-case return	404.80958568559936
Failure rate	0.0

Files

• ppo_humanoid/results.csv: per-episode results for each friction and torque setting.
• ppo_humanoid/summary.json: full benchmark configuration, metrics, and artifact paths.
• ppo_humanoid/leaderboard_row.json: compact leaderboard-ready metrics.
• ppo_humanoid/plots/reward_vs_friction.png: reward aggregated by friction.
• ppo_humanoid/plots/reward_vs_torque.png: reward aggregated by torque scale.
• ppo_humanoid/plots/robustness_heatmap.png: friction x torque robustness heatmap.

Preview

Robustness Heatmap

Reward vs Friction

Reward vs Torque Scale

Intended Use

This dataset is intended for comparing fixed-policy robustness under MuJoCo physics perturbations. It can be used to reproduce leaderboard rows, inspect per-configuration failures, and compare PPO against future RobustHumanoid baselines.

Limitations

The PPO checkpoint used for this run was created as a lightweight baseline/demo. These results should not be interpreted as a tuned or state-of-the-art Humanoid policy. The environment is Humanoid-v4; Gymnasium may warn that a newer Humanoid-v5 environment exists.

Citation

If you use this dataset, please cite or link to the RobustHumanoid benchmark repository and this dataset page.

Source Code

Benchmark implementation and scripts are available on GitHub:

https://github.com/kaus-04/robust-humanoid