Robot and Hand Catalog

Every robot is a RobotConfig instance in configs/robot/, registered by name in that package's __init__.py. build_robot_config(name) imports only that module, so only its assets download.

Hands and rigs

Variants suffixed _side_init and _centered differ only in initial pose or camera framing; _mj variants are the MJCF models used by the MuJoCo evaluator. Not every registered robot has a pipeline: experiments exist for ability, inspire, mano, paxini, sharpa, xhand, the xArm rigs, the G1 rigs, Vega and the MuJoCo Shadow.

What a config carries

Paths to the URDF or MJCF, collision spheres, self-collision ignore list, contact points and contact-avoid points; the base type and initial pose; the link names that name the palm, fingertips, thumb and middle finger; and robokit's EE_COORD_SPEC as ee_coord_spec, with the link it is expressed in as ee_coord_link_name. grasp_robot.hand_coord_spec turns it into palm-forward and fingers-up axes in the hand frame (a gripper's jaws are its palm). Those axes drive the hand-axis cone filter and the init sampler, so a wrong spec shows up as grasps approaching from the wrong side; tests/test_robot_coord_specs.py checks every config's axes against its fingertips and its sampled inits.

Contacts come either from a contact-points file or from a contact_candidates_config that samples per link; scripts/tools/create_contact_points.py authors the former.

ManiSkill agents

Simulation needs a ManiSkill agent per robot, in mani_skill/agents/. The uid is the robot config name, with _floating_base appended for floating hands — xhand_right_floating_base, vega_sharpa_right. Agents load sim-specific URDFs (floating-base, mimic-free variants) rather than the config's own path, and compose a controller: the six virtual base joints plus a position controller for the fingers. A robot with no hand-written agent (dclaw, orca, paxini, the grippers, every assembled rig) gets one from mani_skill/agents/config_agent.py: register_config_agent builds it from the robot config, writing a mimic-free sim URDF (plus the six virtual base joints for a floating hand) under ~/.graspkit/data/sim_urdf/<revision>. The _mj rigs are MuJoCo-only.

Adding a hand

1. Write configs/robot/<hand>.py: one RobotConfig pointing at the robokit asset symbols.
2. Register it in configs/robot/__init__.py.
3. Add a ManiSkill agent in mani_skill/agents/<hand>.py and import it in mani_skill/__init__.py.
4. Author contact points if the hand needs them.
5. Optionally add it to EES in robokit's assets/robots/assembly/composed.py so it can be mounted on any arm, and to the gallery's hand list.
6. Add an experiment config when you want a pipeline for it.

configs/robot/schunk.py plus mani_skill/agents/schunk.py is the smallest complete example.

Assembled rigs

robokit's assets/robots/assembly/composed.py assembles any of seven bases (YAM, xArm7, xArm6, Franka, UR5e, G1, Vega) with any of fourteen end effectors, merging collision spheres and computing the mount transform from the two mount specs, on demand under ~/.robokit/assembly/<revision>. utils/assembly_utils.py adds each base's scene placement and rest pose; its configs are what the gallery's arm dropdown offers. A base+ee name is always assembled, even where a hand-written config shares it. uv run python -m graspkit.gallery.grasp_optim --all --gpus ... optimizes every gallery hand floating and jointly on every base, sim-evaluates all of them, and prints the success table.