Skip to content

long_readme.md

Latest commit

 

History

History
58 lines (39 loc) · 1.39 KB

long_readme.md

File metadata and controls

58 lines (39 loc) · 1.39 KB
Raw

Sim-Evals Setup

Prerequisites

  1. Docker + NVIDIA Container Toolkit
  2. Test: docker run --rm --gpus all nvidia/cuda:11.8.0-runtime-ubuntu22.04 nvidia-smi

Setup

# Make scripts executable
chmod +x dev/*.sh

# Start container (builds on first run)
./dev/up.sh

# Enter container  
./dev/exec.sh


export PYTHONPATH="/workspace/IsaacLab/source/isaaclab:/workspace/IsaacLab/source/isaaclab_assets:/workspace/IsaacLab/source/isaaclab_tasks:${PYTHONPATH:-}"

.venv/bin/python -c "import isaaclab; print(isaaclab.__file__)"
## should say workspace/IsaacLab/source/

.venv/bin/python run_eval.py --headless --scene 4 --visualize-trajectories

Run Evaluation

# Basic
.venv/bin/python run_eval.py --headless --scene 4

# With trajectory visualization
.venv/bin/python run_eval.py --headless --scene 4 --visualize-trajectories
# Then open http://localhost:8888

Policy Server (Host Machine)

Start the policy server on your host before running evals:

CUDA_VISIBLE_DEVICES=1 XLA_PYTHON_CLIENT_MEM_FRACTION=0.9 \
uv run scripts/serve_policy.py \
  --num-samples 4 \
  --temperature 0.5 \
  policy:checkpoint \
  --policy.config=pi0_fast_droid_jointpos \
  --policy.dir=s3://openpi-assets-simeval/pi0_fast_droid_jointpos

Editable IsaacLab

IsaacLab/ source is automatically used (via PYTHONPATH). Modify any file and changes apply immediately!

See QUICK_START.md for details.