ssuleiman/simrig-vision-cartpole

SimRig Vision CartPole

This is the trained reference checkpoint for SimRig's examples/vision_cartpole.py environment. The deterministic actor balances a MuJoCo CartPole from rendered pixels; simulator state is available only to the critic during PPO training.

Observation and action contract

• Actor: 64x64 grayscale images, three-frame stack, plus previous action.
• Critic during training: previous action plus privileged CartPole state.
• Action: one normalized cart actuator command.
• Episode horizon: 1,000 policy steps.
• Renderer: MuJoCo MJX camera rendering through MuJoCo Warp.

Training

• SimRig source: Su1eym4n/simrig commit adc3551.
• PPO steps: 5,079,040 (requested: 5,000,000).
• Vectorized environments: 1,024.
• Training seed: 0.
• Hardware: NVIDIA A100-SXM4-40GB.
• Python: 3.11.15.
• JAX: 0.10.2; Brax: 0.14.2; MuJoCo: 3.10.0; Playground: 0.2.0; Warp: 1.13.0.

The exact resolved configuration and final trainer metrics are included as config.json and final_metrics.json.

Independent evaluation

Deterministic 1,000-step rollouts were run in the recorded training runtime.

Seed	Survival	Terminated	Total reward
0	1,000	No	98.0612
1	1,000	No	98.3289
2	1,000	No	98.1182
3	1,000	No	97.6037
4	1,000	No	98.1110

Survival mean/min/max: 1,000/1,000/1,000. Termination rate: 0%. Mean total reward: 98.0446. These results establish full-horizon balancing across the tested resets; they do not yet constitute a camera-ablation robustness test.

Run with SimRig

Clone SimRig and install the Playground and Hugging Face extras in a Python 3.11 environment. Native pixel rollout requires a JAX-visible CUDA GPU and MuJoCo Warp.

git clone https://github.com/Su1eym4n/simrig.git
cd simrig
git checkout adc3551
python3.11 -m venv .venv
.venv/bin/python -m pip install -e ".[playground,hf]"

.venv/bin/simrig eval \
  hf://ssuleiman/simrig-vision-cartpole/policy.params \
  --env examples/vision_cartpole.py \
  --hf-revision v1 \
  --steps 1000 \
  --seed 0

.venv/bin/simrig preview \
  hf://ssuleiman/simrig-vision-cartpole/policy.params \
  --env examples/vision_cartpole.py \
  --hf-revision v1 \
  --auto-reset \
  --port 8765

SimRig downloads policy.params together with its sibling config.json so it can reconstruct the exact vision CNN. Runtime mismatches are rejected by default. --allow-runtime-mismatch is appropriate only for explicitly qualitative preview checks.

Files

• policy.params: exported Brax policy parameters.
• config.json: resolved training, network, vision, runtime, and source data.
• final_metrics.json: final PPO trainer metrics.
• evaluation.json: independent five-seed rollout results.
• vision_cartpole.py: environment source snapshot used for training.