5hadytru/so101_bench_InternW0-Delta

SO-101 Bench: InternW0-Δ full fine-tune

InternW0-Δ (video expert from Wan2.2-TI2V-5B + 1.2B action expert, RynnBrain1.1-2B VLM) post-trained in full from InternRobotics/InternW0-Delta-Base on the SO-101 Bench simulated teleop dataset 5hadytru/so101_bench_sim (3,629 episodes, 1,996,120 frames; overhead + wrist cameras).

• Official checkpoint (repo root): step 18200, the final step of a 36 h run on 4x A100-80GB (global batch 128, AdamW 5e-5 cosine, ~1.17 epochs). Validation (39-episode SO-101 Bench val set): 20/39 (51.3%).
• step016000/: the best validated checkpoint, 22/39 (56.4%), kept as a validation-selected alternative.
• Files: model.pt (InternW0-Delta weights), config.yaml (the run's resolved Hydra config), dataset_stats.json (normalization).
• Inputs: overhead above wrist on the model's 384x256 canvas (192x256 each); frame-conditioning memory (episode's first frame + the frame 32 steps back); 6-D absolute joint targets in LeRobot units; 32-action chunks at 30 Hz, executed 16 per replan.
• Code: InternRobotics/InternW0-Delta @ 90801ba; recipe, configs and server in the SO-101 Bench repo (scripts/InternW0_fine_tuning/, scripts/internw0_server.py).