Fast-GraspD – Differentiable Grasp Optimisation

Project | Paper

Fast-GraspD searches for stable grasps on a given object mesh using a fully-differentiable Warp simulation. Only the Allegro right hand is supported at the moment.

---

Install

pip install -r requirements.txt  # tested with warp-lang==1.7.1, python 3.12

---

Run

python scripts/collect_grasps_allegro.py \
  collector_config.obj_set=ycb \
  collector_config.obj_name=003_cracker_box

All parameters live in conf/collect_grasps/config.yaml and can be overridden from the command line via Hydra (collector_config.*). Tune the loss weights (w_*_loss) for best results.

---

Outputs

<output_dir>/<hand_name>_<obj_set>/<obj_name>/
    *.npy              # final joint angles
    *.usd              # final grasp (optional)
    *_opt_traj.usd     # optimisation trajectory (optional)

USD export is toggled by render_final_grasp and render_opt_traj in the config.

---

TODO

• Add support for Barrett hand
• Add support for Shadow hand
• Set tested default configuration values