pi05_mvtoken_22_27_04

pi0.5 SFT checkpoints fine-tuned on the MVTOKEN_22_27_04 real-robot dataset.

Checkpoints

Two checkpoints from the same run, laid out as in the training output directory:

global_step_2000/actor/model_state_dict/full_weights.pt   # train loss ~0.014,  93 passes
global_step_4000/actor/model_state_dict/full_weights.pt   # train loss ~0.008, 177 passes
norm_stats.json                                           # shared by both

Evaluate both. The dataset is small (5,070 frames), so by step 4000 every frame has been seen ~177 times. For comparison, the SweepIntoDustpan-v1_Real run reached the same loss level after only ~15 passes — a lower training loss here does not imply a better policy.

norm_stats.json is the same file as in aaroncaozj/pi05_norm_stats_collection → MVTOKEN_22_27_04/.

FSDP distributed-checkpoint shards (dcp_checkpoint/*.distcp, ~22 GB each) are not included; they only serve to resume training and are not needed for inference.

Data

episodes	101
frames	5,070
fps	4
duration / episode	~12.5 s
state	7-D end-effector pose + gripper width
actions	7-D delta EE pose + binary gripper, ~2 cm per step
cameras	image (agentview) + wrist_image; back_image is an all-zero placeholder and is masked out per sample by FrankaEEInputs

Training

Trained with RLinf (examples/sft/config/mvtoken_sft_openpi.yaml), openpi data config pi05_mvtoken.

base model	pi05_base
action_horizon	4
hardware	7 x H200 NVL
micro / global batch	32 / 224 (grad accum 1)
optimizer	AdamW, lr 2.5e-5, cosine, 1000 warmup