RACE2 on VLABench — checkpoints and data
Artifacts of the RACE2 (transition-aware fine-tuning of pi0.5) experiments on the VLABench primitive tasks. Code: Seonghoon-Yu/VLABench.
Data
| folder | contents |
|---|---|
data/vlabench_phase_5000/ |
5,000 freshly generated expert demonstrations (500 per task, LeRobot v2.0, same schema as SeonghoonYu/vlabench_primitive_ft_lerobot_224) with generator-logged transition labels: labels/vlabench_phase_cps.npz (reference phase boundaries, 4.49/ep) and labels/vlabench_phase_pelt.npz (PELT labels on the same episodes, 3.81/ep). See data/vlabench_phase_5000/README.md. |
The original training set (5,000 episodes, PELT labels) is SeonghoonYu/vlabench_primitive_ft_lerobot_224.
Checkpoints
| folder | run |
|---|---|
teacher_official/ |
pi0.5 VLABench teacher (official) |
plainft_h{10,15,20,25,30}_40k/ |
PlainFT baseline, action horizon H, 40k steps |
h{10,15,20,25,30,35}_40k/ |
RACE2, action horizon H, 40k steps |
acot_h{10,20}_40k/ |
ACoT-VLA reproduction (JAX/orbax checkpoint; see its README) |
ablation/abl_*_h20_40k/ |
RACE2 ablations at H20 (actionhead, nojitter, notf, speedcp, timing_only, vlmhead, …) |
h25_support/, h30_h35_support/ |
openpi overlays (configs, CPS assets, serving entrypoint) needed to load the H25 / H30–H35 checkpoints |
Each run folder carries its assets/ (norm stats), weights, and where available README.md,
training_config.json, provenance.json and logs/.
Download
from huggingface_hub import snapshot_download
snapshot_download("SeonghoonYu/RACE2_vlabench", allow_patterns="data/vlabench_phase_5000/*",
local_dir="/path/to/lerobot/SeonghoonYu") # -> .../SeonghoonYu/data/vlabench_phase_5000