AgenticWAM RoboTwin models

Both models use three cameras and 14-D absolute joint actions. They share dataset_stats.json; task_skill_prompt.json supplies the WAM instruction prompts. Evaluation settings are provided in wam.yaml and hybrid.yaml and in configs/robotwin/eval/ in the code repository.

Setup

Follow installation and the RoboTwin guide. From the code repository root:

hf download miniFranka/agentic-wam-posttrain-robotwin-suit --local-dir checkpoints/robotwin_release
cp configs/robotwin/site.example.yaml configs/robotwin/site.yaml

Configure your machine paths in site.yaml and configs/paths/default.yaml. Public model components are Wan-AI/Wan2.2-TI2V-5B (Wan2.2_VAE.pth), nvidia/Cosmos-Reason1-7B, JetLM/SDAR-1.7B-Chat, and Qwen/Qwen3-VL-2B-Instruct. Prepare RoboTwin, cuRobo, Vulkan, and the robot/object assets. Hybrid additionally uses Molmo, SAM3, and Contact-GraspNet; their setup is in codegen/sim/SETUP.md.

Evaluate

Run one episode per clean/randomized setting:

python scripts/eval_robotwin_release.py wam --tasks click_alarmclock --plan smoke --gpus 0
python scripts/eval_robotwin_release.py hybrid --tasks click_alarmclock --plan smoke --gpus 0 --vis-gpu 1

Remove --tasks click_alarmclock --plan smoke for the full 50-task benchmark. Hybrid starts or checks perception on --vis-gpu; keep policy and perception GPUs separate for multi-episode runs. Add --dry-run to inspect the plan or --render to save videos.

The commands select the matching weights and evaluation config automatically. Outputs and resumable reports are under evaluate_results/. The hybrid task programs are included in configs/robotwin/tasks_indep/.

Both models passed clean/randomized smoke evaluation on click_alarmclock with one episode per setting. manifest.json records checksums and validation details.