v8-flow_h16_s8

Flow-matching policy (horizon 16, n_action_steps 8), 140K steps, batch 24, EMA 0.9999. Baseline for v8-flow_h16_s8_stagelabel: same data with the stage input observation.environment_state dropped.

  • Root = step 140000 (final). checkpoints/100000/, checkpoints/120000/ = earlier snapshots. train.log = full log.
  • Code: lerobot git commit 2af91da5 (branch diffusion-flow-react), trained on lab PC (RTX 3080 Ti), 2026-11-03 02:55โ€“14:02.
  • Data: wwqcleverghost/v8-handover2-stagelabel revision d299c93fc66b0c62014e70916c5d304be7bfca67 (45 eps, 23,328 frames, 30 fps), 12-dim arm-only.
  • Inputs: observation.state, left_wrist, right_wrist, head(RGDB). W&B: wwqcleverghost/handover run ly72tvb5.
Downloads last month
2
Safetensors
Model size
0.3B params
Tensor type
F32
ยท
Video Preview
loading