matCercola18/so101_gpl_data

SO-101 campaign data: built datasets, plans, planner runs

Derived from the raw demos matCercola18/Toilet_paper_SO101, Two_cubes_SO101 and Clean_Workspace_SO101. The code is in the GitHub repo latent-reasoning-vla (branch so101-hardware, hardware/). The layout mirrors data/so101/, so the scripts' default paths work after huggingface-cli download matCercola18/so101_gpl_data --repo-type dataset --local-dir <Continuous-Reasoning>/data/so101.

• so101_gpl_v1/: LeRobot v3 dataset, 501 cleaned episodes (TP 139, TC 125, CW 237), local calibration frame, true quantile statistics; hardware/build_dataset.py --version v1. Splits and index: hardware/labels/so101_gpl_v1_*.
• so101_gpl_v0/: the first build (249 episodes).
• sketches_v{0,1}.npz/.json: per-episode contact-sketch plans (sketch_abs = executor input, sketch_rel = reasoner target), hardware/so101_sketch.py; sketches_v0_prev is the previous v0 build.
• runs/cond_*: reasoner conditions (hardware/so101_cond.py). cond_s0full, cond_s0expert: last prefix token of the S0 executors (v0). cond_v1_siglip: pi05_base SigLIP features (v1), used by reasoner_v1.
• runs/reasoner_v0, runs/reasoner_v1: hardware/so101_reasoner.py runs (3 seeds, results, logs). reasoner_v1/seed0.pt is the robot's reasoner; its scene encoder (siglip.safetensors) is in matCercola18/so101_gpl_v1_j_full/reasoner/.
• runs/s1_quick: the first plan adapter on the frozen S0. It ignored the plan (O2 switch 0.009), which led to joint training.
• raw/*_meta_*_backup/: the raw datasets' metadata before they were extended in place on 2026-10-01.
• *_build.log: build logs.

Checkpoints: matCercola18/so101_gpl_v1_j_full (robot bundle), so101_gpl_v1_j_expert, so101_gpl_v0_j_full, so101_gpl_v0_j_expert, so101_gpl_v1_f_base, so101_gpl_s0_full, so101_gpl_s0_expert, so101_gpl_v0_s1c. Robot-session logs: matCercola18/so101_gpl_eval.