cosmos3-edge-solar-panel-225-20k

Cosmos3 Edge — SO101 solar-panel cleaning

Training launched on 2026-09-15. The initial step-10 checkpoint was uploaded and verified, and the run resumed from its training state toward step 20,000. See latest_checkpoint.json for the latest verified checkpoint. Target: 20,000 steps on 8 H100 GPUs. Completed checkpoints will appear under checkpoints/, with a verified latest_checkpoint.json pointer.

Inputs: current wrist image and task instruction; no observation history or robot state. Outputs: 16 frame-wise relative end-effector pose actions and gripper values at 15 Hz. Task: Use the brush to clean the solar panel on top of the helicopter.

Dataset: 225 source episodes, split into 202 training and 23 validation episodes, seed 42. The processed H.264 dataset is private. No training videos or episode parquet files are published here.

Base LR 5e-5, action-head multiplier 5, 500-step warmup, 20,000-step linear schedule, effective batch 64 (8 samples × 8 GPUs), BF16 and full activation checkpointing. LIBERO-derived action policy adapted to Cosmos3 Edge and SO101; this is not an official NVIDIA checkpoint.

Code and exact recipe: https://github.com/Rebis-IvLabs/cosmos-framework/tree/codex/solar-panel-edge-20k/examples/solar_edge Latest NVIDIA upstream merged: 3f94d25d564ff0a982ac3e540cb83bfe4a5e5a11.