Bariona/fact-world-action-model

FACT — Failure-Aware Causal Training for World-Action Models

Interactive demo of `Bariona/fact-wam`, the RoboTwin 2.0 checkpoint from FACT (project page, code).

FACT is a single causal diffusion transformer (fine-tuned from Wan2.2-TI2V-5B) that jointly denoises the next 48 robot actions, the future video and a task-progress value. The demo exploits the pipeline's gt_action_condition path to imagine two futures from one observation — one under FACT's own plan, one under an action you deliberately sabotaged (freeze the arms, reverse the motion, open the grippers, swap the arms, …) — and compares the predicted task-progress / failure value of each.

Both rollouts share the same reference frames, instruction, seed and noise; only the action conditioning differs. Inference follows the released evaluation recipe: 384×192 three-view canvas, 48-action chunk, 5 keyframes, 20 UniPC steps, flow shift 3.0, guidance 0.0, bf16.

Credits

• —world_action_model/ is copied unmodified from Bariona/FACT (Apache-2.0, see LICENSE).
• —The example observations in examples/ are real frames from `Bariona/robotwin-v2` (MIT), generated with RoboTwin 2.0.