ForceSight — pi0.5 fine-tuned checkpoints
Fine-tuned pi0.5 (openpi) policies for ForceSight manipulation tasks. All checkpoints are from step 20000, trained on 6/3 data.
Checkpoints
| Folder | Variant | Description |
|---|---|---|
pi05_6_3/ |
Baseline pi0.5 | Vanilla pi0.5 fine-tune, no tactile. |
encoder_6_3/ |
pi0.5 + tactile encoder | Conv-Based encoder for tactile images |
tactile_6_3/ |
pi0.5 + tactile | Tactile images are augmented as camera inputs to the VLA model |
tapvla_6_3/ |
pi0.5 + annotation | Tactile sensor data is annotated directly on the VLA images |
Each folder contains params/ (orbax weights) and assets/ (normalization stats — required for inference). |
Setup
- Base model: pi0.5 (openpi)
- Robot: Franka Emika Panda + Franka Hand
- Tasks: Medicine, Balance, Gear Insertion, Plug Insertion.
- Training: 20000 steps, 4 A6000 GPUs.
Loading
Download a single checkpoint:
hf download mlshehab/forcesight --include "pi05_6_3/*" --local-dir ./forcesight
Load with openpi:
from openpi.policies import policy_config
from openpi.training import config
cfg = config.get_config("<FILL IN: config name, e.g. pi05_forcesight>")
policy = policy_config.create_trained_policy(cfg, "./forcesight/pi05_6_3")
Note: the tactile and TAP-VLA variants require a custom openpi config/fork. See openpi.