GitHub

Distill Claude's per-frame action choices into a fast feedforward BC policy.

See CLAUDE.md for the full pipeline and sibling-repo coupling.

Quickstart

# 1. Build the labeled dataset (offline, hours)
python scripts/build_bc_dataset.py \
    --wm-checkpoint C:/Projects/canvas-autonomous-learner/runs/red_kong_1b/checkpoints/ft_20260504_145349/best.pth \
    --corpus-manifest C:/Projects/canvas-autonomous-learner/runs/red_kong_1b/canvas/merged_20260504_145349/dataset_meta.json \
    --hf-cache-root C:/Users/irvin/.cache/huggingface/lerobot \
    --num-samples 5000 \
    --parallelism 4 \
    --output ./datasets/bc_red_kong \
    --balance-by-action \
    --dump-contact-sheet 50
# 2. Train the BC policy (minutes)
python scripts/train_bc.py \
    --dataset ./datasets/bc_red_kong \
    --output ./checkpoints/bc_red_kong
# 3. Eval (confusion matrix + contact sheet)
python scripts/eval_bc.py \
    --checkpoint ./checkpoints/bc_red_kong/best.pt \
    --dataset ./datasets/bc_red_kong

Read the original on github.com ↗