GM717/instinct-wam-robotwin-clean

InstinctWAM (RoboTwin 2.0)

A compact World-Action Model (video world-model + action expert) for bimanual manipulation on RoboTwin 2.0 (aloha-agilex). Trained from scratch using only permissively-licensed inputs:

• WAN-2.2-TI2V-5B (teacher / structured-slice init) — Apache-2.0
• RoboTwin 2.0 trajectories — Apache-2.0
• No third-party trained checkpoints were used (see NOTICE).

Architecture

• Compact video DiT (dim 2048, 12 layers, 16 heads, head_dim 128; ~849M) distilled from WAN-2.2 via structured layer slicing; future-latent size 192×160 (RT variant).
• Action expert (dim 768, 12 layers, chunk 16, state/action dim 14; ~139M), joint attention (MoT).
• Flow-matching / rectified-flow objective; asymmetric denoising at inference (RT: 2 video-refresh steps).

Training

Curated 16 hard/long-horizon RoboTwin tasks → 9,254 episodes → 2.55M samples. Three-stage recipe, 80k steps/stage on 2×H100:

1. Compact video distillation (hidden+motion KD in 256-d PCA space + GT flow-matching)
2. Frozen-video action training
3. Joint video-action refinement

Evaluation (16 tasks × 20 episodes, success rate %)

budget	demo_clean	demo_randomized
80k×3 (this model)	61.6	58.4

Files

• instinctwam_robotwin2.pt — the deployable weights (final, joint-refined).
• action_stats.json — qpos mean/std normalization (required at inference).
• NOTICE — third-party attribution (WAN-2.2, RoboTwin 2.0).

Usage

Code / architecture: https://github.com/william-Dic/instinctwam_clean

from instinct_wam import InstinctWAM
model = InstinctWAM.from_pretrained("GM717/instinct-wam-robotwin-clean")  # pulls instinctwam_robotwin2.pt

Pair with action_stats.json for action de-normalization. RT inference settings: num_inference_steps 10, video_refresh_steps [0, 1], future_video_size [192, 160].

License: Apache-2.0 (see NOTICE for third-party components).