half-cheetah-local-sac

SAC HalfCheetah-v5

This is a Stable Baselines3 SAC policy trained locally on Gymnasium HalfCheetah-v5.

The environment uses a small anti-flip reward guard. It penalizes extreme torso plane angle and low torso height, then terminates clear fall or belly-slide exploit postures.

Files

• sac_half_cheetah.zip: Stable Baselines3 SAC checkpoint.
• videos/initial.mp4: random policy before training.
• videos/final.mp4: trained policy rollout.
• sac_cheetah/: minimal training, environment, check, and video code.
• pyproject.toml: Python dependency setup.

Training

• Algorithm: Soft Actor-Critic
• Implementation: Stable Baselines3
• Policy: MlpPolicy
• Environment: HalfCheetah-v5
• Timesteps: 300000
• Seed: 7
• Device: Quadro P520 with torch==2.7.1+cu118

Evaluation

Single deterministic rollout with seed 8.

These numbers are a local smoke evaluation, not a benchmark sweep.

Load

from stable_baselines3 import SAC

model = SAC.load("sac_half_cheetah.zip", device="auto")

Use the wrapper in sac_cheetah.envs if you want evaluation to match this model card.

from sac_cheetah.config import TrainConfig
from sac_cheetah.envs import make_env

cfg = TrainConfig()
env = make_env(cfg.env_id, cfg.seed + 1, render_mode="rgb_array")

Limitations

This policy is only tested on Gymnasium HalfCheetah-v5 with the included anti-flip wrapper. It is not meant for real robots, safety-critical systems, or transfer to other MuJoCo tasks without retraining.