UR5E-LeapHand-Tactile
Code, configurations, simulation assets, trained policies and evaluation summaries for visuotactile dexterous grasping experiments run in the Genesis simulator with GelSight-style tactile sensing (UR5e arm; DG-3F-M, Allegro and LEAP hands). Everything here was exported from the training server on 2026-11-10.
Per-episode training and evaluation records are in the companion dataset repo Linpeng502502/UR5E-LeapHand-Tactile-runs.
Layout
| Path | Contents |
|---|---|
code/mh_z5090/ |
Versioned release code (releases/, each with its SHA and RELEASE.json), job scripts, logs and completion markers |
code/tactile-genesis/ |
Genesis tactile (GelSight) simulation code |
code/genesis-world-main/ |
The Genesis source tree used on the server (Apache-2.0, see its LICENSE); may contain local changes |
project/ |
Project root as mounted at run time: configs, pars_grasp, runtime_assets (arm and hand models), pyproject.toml |
project/outputs/multihand_framework_20261026/ |
Pilot inputs, sensor models (gel1–gel3o), coarse grasp specs, offline Hertz contact diagnosis |
models/<run path>/ |
Top-level files of each finished training run: policy_*.npz, ckpt_*.pt, training_plan.json, training_state.json, train_log.jsonl |
results/<evaluation path>/ |
Evaluation summaries: results.json, results.md, eval_plan.json, eval.log |
env/ |
Package lists and platform information for rebuilding the Python environments |
UPLOAD_MANIFEST.md |
The upload plan (Chinese): what was uploaded, from where, and what was left out |
Run paths under models/ and results/ are relative to the project's outputs/ directory, so they match the archive
paths in the dataset repo. Extracting all dataset archives into one directory rebuilds the outputs/ tree.
License
No license has been chosen yet. Third-party code keeps its own license (for example code/genesis-world-main/LICENSE).