Downloads · 30 days
0
isldgist/pi05-panda-weightgen-two-teacher-inits-ee6
pi05-panda-weightgen-two-teacher-inits-ee6 is a robotics model from isldgist. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for jax. The card lists the license as apache-2.0.
This repository contains two task-agnostic Panda zero-shot checkpoints. Both checkpoints use the same Sawyer + IIWA + UR5e joint checkpoint at update 19999 as the final base model and generate rank-8 updates for the c…
Downloads · 30 days
0
Access
Public
Updated Sep 14, 2026
Repo size
21.8 GB
Likes
0
Public
Click a slice to open those files.
Other24.9 GB · 100%
From the Hugging Face model README
This repository contains two task-agnostic Panda zero-shot checkpoints. Both checkpoints use the same Sawyer + IIWA + UR5e joint checkpoint at update 19999 as the final base model and generate rank-8 updates for the complete pi0.5 action decoder. They differ only in how the HyperNetwork was initialized before the 10,000-update behavioral zero-shot meta-training stage.
| Repository path | HyperNetwork teacher initialization | Teacher updates |
|---|---|---|
joint19999_teacher_init_2k/panda | Sawyer + IIWA + UR5e joint checkpoint, update 19999 | 2,000 |
official_pi05_base_teacher_init_5k/panda | Official OpenPI pi05_base checkpoint | 5,000 |
The final checkpoint formula in both cases is:
W_panda = W_joint_19999 + Delta_W_panda
Delta_W_panda is produced from the Panda VLM context and 49-D simulator
capability vector. The generator predicts rank-8 factors for the action-expert
attention, MLP, AdaRMS, action-input/output, and time-MLP matrices, plus direct
bias deltas where applicable. The vision and prefix/text encoders are inherited
unchanged from the joint checkpoint.
Both checkpoints are task-agnostic and are shared by square, threading,
stack, and stack_three. They consume a 6-D end-effector pose state and emit
the unchanged 7-D OSC_POSE action. Both retain the joint checkpoint's shared
normalization statistics.
Each checkpoint directory contains inference params, normalization assets,
complete.json, and weightgen_metadata.json. Optimizer state is not included.