OpenWAM IDM: CF64k
Project | Code | Inference guide
Robot control learned from varied physical motion. OpenWAM IDM is an inverse dynamics model that predicts the actions connecting a current observation to a desired visual future. Paired with a video predictor, it turns an imagined motion into robot controls.
This checkpoint uses 100% counterfactual dynamics during IDM post-training on the LIBERO-Long (LIBERO-10) CF64k collection. These examples explore different action sequences from shared starting states, extending the model's experience beyond demonstrated task trajectories. The model builds on OpenWAM's robot-video-pretrained backbone and is trained for 20,000 post-training steps.
The IDM reads the current agent and wrist camera views, robot proprioception, and a short future video represented as latents. It predicts end-effector delta controls and gripper commands, without task-language conditioning or past actions. This local formulation lets the frozen IDM work alongside task-adapted video predictors without retraining the action component.
Evaluation
| Evaluation | Success |
|---|---|
| LIBERO-10, 500 episodes | 93.8% |
| Transfer to four LIBERO-90 tasks, 50 episodes each | 93.5% |
Results measure the composed video-to-action system with this IDM frozen. LIBERO-10 uses the OpenWAM VTA step-9700 video predictor (469/500 successes). Transfer uses task-tuned video predictors on tasks 21, 45, 64, and 74; the reported score is the mean across those four tasks.
Use with OpenWAM
Pair this checkpoint with a video predictor through OpenWAM's video-to-action pipeline. The release includes model weights and an inference configuration; the guide covers setup, shared frontend assets, and evaluation.
For the counterpart trained with both counterfactual data and demonstrations, see OpenWAM IDM Mixed CF64k.
License
Model weights are released under AGPL-3.0-only. Upstream attribution and license terms are retained in NOTICE and LICENSE.upstream.