Downloads · 30 days
0
jellyho/pi05_yam_lego_taxi_alphaflow
pi05_yam_lego_taxi_alphaflow is a robotics model from jellyho. Use it for the robotics task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
pi0.5 fine-tuned with the AlphaFlow curriculum (flow matching - mean-velocity MeanFlow), so the action expert samples a full action chunk in one forward pass (numsteps=1) instead of the usual 10-step ODE. Trained on j…
Downloads · 30 days
0
Access
Public
Updated Aug 26, 2026
Repo size
74.6 GB
Likes
0
Public
Click a slice to open those files.
Other74.6 GB · 100%
From the Hugging Face model README
pi0.5 fine-tuned with the AlphaFlow curriculum (flow matching -> mean-velocity
MeanFlow), so the action expert samples a full action chunk in one forward pass
(num_steps=1) instead of the usual 10-step ODE. Trained on jellyho/yam_lego_taxi_s300
for 200k steps.
Provenance correction (2026-08-24): this v1 run trained on ALL 347 episodes (300 success + 47 failure -- success_only was not set) and with 08-03-era norm stats whose action q99 is ~37% narrower than the training data (asset
jellyho/yam_lego_taxi, md5 60cb2ee2). The model is internally consistent (its checkpoint carries the same stats it trained with), but comparisons against success-only baselines are confounded, and robot-space (per-policy unnormalized) evaluation is the only fair protocol. A corrected v2 (success-only 300 eps + the BC-h30 baseline's exact stats, provenance-stamped) is training and will be published separately.
src/openpi/models/pi0_alphaflow.py on the integration branch of
jellyho/ACRFT (openpi fork). Config: pi05_yam_lego_taxi_alphaflow.c4vy84yy (delta^2 0.052 -> 0.0026, no instabilities).50000/-150000/ -- curriculum-phase checkpoints (alpha still annealing).200000/ -- recommended: end of the official curriculum (discrete floor alpha=5e-3).
The 1-step gate numbers above are from this checkpoint.250000/, 300000/ -- pure-JVP MeanFlow continuation (alpha pinned to 0), same wandb run.WARNING (2026-08-26): the JVP-continuation checkpoints (250000, 300000) produce visibly unstable, noisy rollouts. Cause (wandb c4vy84yy): after the alpha=0 JVP switch the network's directional derivative blew up (dudt_absmax mean ~55, max 294), the target clamp saturated at 4.0 continuously, and delta^2 jumped 65x (0.003 -> 0.19) and never recovered -- the JVP objective never converged in bf16 on this architecture. Use 200000 for deployment and experiments.
200000/ holds an openpi/orbax checkpoint: params/ (no optimizer state) and
assets/ (norm stats). Load with openpi:
from openpi.policies import policy_config
from openpi.training import config as tconfig
cfg = tconfig.get_config("pi05_yam_lego_taxi_alphaflow")
policy = policy_config.create_trained_policy(
cfg, "<local download>/200000", sample_kwargs={"num_steps": 1}
)
N-candidate sampling for critic-side best-of-N is available via
model.sample_n_actions(...) / scripts/serve_patch_critic.py --flow-steps 1.