Piper Pick-Place LeRobot RGB+tactile Dataset
Current repository ID: Alexxz3/piper_pick_place_4obj_pi05_tactile_400ep_v1.
Purpose
This dataset/artifact repository contains canonical LeRobot RGB+tactile training dataset for the Alexxz3 Piper pick-and-place release.
Release context
The release supports reproducible simulation evaluation, LeRobot-compatible training data, pi0.5/ACT policy work, RGB and RGB+tactile modalities, and future real-world deployment on the AgileX Piper arm. Keep/delete decisions are documented in docs/release/hf_asset_inventory.md. Keep the raw dataset, LeRobot RGB+tactile dataset, LeRobot RGB-only dataset, and final model repos under Alexxz3; helper asset and source-archive repos are not required.
Dataset schema
Raw RoboTwin rollouts preserve simulator episode structure. LeRobot datasets expose RGB camera observations, robot proprioception, actions, and, for RGB+tactile, tactile arrays and masks from the fingertip sensors. RGB-only remains available because it is a smaller compatibility dataset with modality-specific normalization and metadata, even though RGB+tactile contains the richer sensor stream.
Evaluation linkage
Final model-card results use the randomized in-distribution eval_id suite: bottle, can, toy car, and shampoo with 25 episodes per task. See docs/piper_id_eval_report.md for the full table and analysis.
Limitations
This data is simulation data and does not certify real-world behavior. Tactile deployment requires matching sensor layout, calibration, masking, and normalization statistics.
License
Released under Apache-2.0 unless upstream RoboTwin, LeRobot, OpenPI/pi0.5, ACT, or object-asset dependencies impose additional obligations.
- Downloads last month
- 92