You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

SAMBAR π0.5 LIBERO checkpoints

Pretraining and finetuning checkpoints for SAMBAR.

Pretraining checkpoints

Pretraining curriculum	Checkpoint directory
Narrow, 32 tasks	libero-32/15000
Broad, 117 tasks (27 + 90)	libero-117/9999

Each numbered directory contains Orbax parameters under params/ and normalization statistics under assets/. The surrounding run folder includes the original hyperparameters and any logs provided with that run.

To download the 32-task checkpoint after obtaining model access, run:

hf download Berkeley-ICON-Lab/SAMBAR-checkpoints --include 'libero-32/15000/*' --local-dir ./checkpoints

For the 117-task checkpoint, use --include 'libero-117/9999/*'. Use the numbered directory as the pretrained checkpoint path. For sequential SAMBAR training, compute the initial LIBERO importance from the matching pretraining demonstrations using the importance export instructions.

Finetuning checkpoints

These runs start from the 32-task LIBERO checkpoint. The sequential checkpoints are the final stage after tasks 3 → 8 → 11. Each independent run finetunes on one held-out task. The links point to the saved Orbax step directories.

Experiment	SAMBAR	SFT
Sequential, tasks 3 → 8 → 11	9999	9999
Independent, task 3 (pot-on-stove)	19999	9999
Independent, task 8 (bowl-in-drawer)	19999	9999

Downloads last month: -; Downloads are not tracked for this model. How to track

Video Preview

Robotics