curobo

Collision-aware cuRobo motion planning as in-process tools. Trajectories in/out are gap Trajectory dicts (waypoints: [{positions: float64[dof]}]); worlds are gap WorldConfig dicts (build one with geometry.build_world_config).

When to use

• Tabletop pick: geometry.top_down_grasp_candidates → curobo.plan_to_grasp_poses (pass the whole fan as the goalset; check goalset_index for which one was reached).
• Transport after grasping: curobo.plan_with_grasped_object with the object's mesh name from build_world_config.
• Drawers/doors: curobo.plan_grasp_motion (approach → grasp → lift/pull with gripper commands interleaved), or curobo.plan_directed_linear for a pull along one axis with orientation locked.

Install

cuRobo JIT-compiles CUDA extensions at install time — build isolation must be off and CUDA_HOME must point at a toolkit matching your torch build:

export CUDA_HOME=/usr/local/cuda     # toolkit matching torch's CUDA version
uv sync --extra curobo               # (pip: pip install -e ".[curobo]" --no-build-isolation)

If the import fails at tool-call time the tools raise a ToolError with this recipe. First planner call per process pays JIT/warmup latency; the MotionGen/planner instances are cached and reused (HyRL pattern — recreating them per call corrupts CUDA graph state).

Gotchas (carried over from the service + curobo_api)

• Frames: grasp/target poses are in the robot-base frame (cuRobo treats the robot base as world origin). With grasp_pose_is_fingertip=True (default) grasp positions are fingertip-pad centers and converted to the panda_hand frame solver-side (offset 0.1029 m along hand Z).
• Ignore the grasp target: pass its mesh name in ignore_obstacle_names for plan_to_grasp_poses / batch_grasp_feasibility — closing on the target is not a collision.
• robot_collision_sphere_buffer default −0.01 shrinks robot collision spheres 1 cm; reduces IK_FAIL against dense perception meshes. Negative is intentional.
• GPU access is serialised by a module lock (cuRobo is not thread-safe); CUDA/"graph capture" errors invalidate the cached planners automatically before raising PlanningFailed.
• use_cuda_graph must stay False for the validators (check_start_state requirement) and for varying world/start setups.
• curobo version split: plan_to_grasp_poses, plan_grasp_motion, plan_directed_linear, plan_linear, plan_to_pose, plan_with_grasped_object target curobo v0.8 (MotionPlanner API); solve_ik and batch_grasp_feasibility are built on the v0.7 IKSolver API that v0.8 removed — on a v0.8-only install they raise PlanningFailed ("not supported on cuRobo v0.8"); plan via the goalset tools instead. The validators use the v0.7 MotionGen path too.
• validate_joint_trajectory_grasped always invalidates the planner cache afterward so the attachment cannot leak; expect the next planning call to re-create the planner.
• Planning failures return success=False (with failure_reason where the RPC had one); infrastructure errors raise PlanningFailed / ToolError.
• Set debug_out_dir on the grasp/transport planners to dump world + robot sphere OBJ/PLY debug artifacts on failure (default ./curobo_debug*).