Collision environment

Collision is URDF-driven. The robot’s covering spheres and self-collision ranges are baked into the generated grid.cuh as GRiD’s grid_collision namespace (pass --collision to scripts/codegen/generate_grid.py). Hard/both refinement uses the warp-scoped grid_collision::warp::config_free on existing joint transforms; the kernel also scores post-solve via grid_collision::collision_distance (soft penetration cost) and, in hard mode, filters with grid_collision::config_free (see Solver kernel). There is no hand-written per-robot collision header. Supported fixed-base serial arms get FK and collision from one codegen step; see Custom robot for the solver’s model restrictions.

hjcdik.worlds converts MotionBenchMaker JSON into arrays. Solver.set_world validates spheres, boxes and capsules (cylinders become capsules) and creates an owned uploaded world. The native CollisionWorld wraps GRiD’s grid_collision::DeviceEnvironment; its Environment view is passed by value into the solver kernels. See Python API for array shapes.

The default collision="hard" on a collision build checks self-collision and the environment, then excludes colliding candidates. It can return zero solutions. soft only ranks by environment penetration cost and offers no collision-free guarantee; both ranks and strictly filters. The final filter is authoritative. These checks do not certify a motion path or guarantee that the returned candidate reaches the requested pose.

Collision-aware early stopping, seed ranking and repair help search for a free candidate. A free candidate within 5 mm / 0.05 rad can be retained as a fallback if this run finds no exactly converged free candidate. Always inspect both reported errors for the same candidate; an empty result is not a proof that the problem is infeasible.

Each selected problem must explicitly describe its obstacles. Use "obstacles": {} for an empty environment (self-collision is still checked). Supported obstacle keys are sphere, cuboid and cylinder; unknown keys are rejected. Cuboids and cylinders take a pose; spheres take a pose or a bare position. Dimensions, radii, and lengths must be positive finite numbers; poses use meters and [x,y,z,qw,qx,qy,qz] with a nonzero quaternion. Cylinders are conservatively represented as capsules. Guarantees are relative to the generated sphere model, which excludes the fixed Panda base geometry.

World ownership and replacement

Each Solver owns one world bound to its CUDA device. Repeated solves reuse that uploaded world without parsing or comparing JSON. set_world replaces it; clear_world removes the environment while hard mode still checks self-collision. Invalid input or a failed upload leaves the previous world installed. Replacing a world waits for an in-flight solve, so one world applies to its entire target batch; returned results own their host data.

Different solver instances have independent worlds. Native ownership releases device memory on the device that allocated it. collision_margin_m widens environment checks only; self pairs retain zero margin. self_collision=False skips only self pairs. Both options apply to hard/both modes, not the soft penetration cost. This does not establish safe motion between configurations or coverage of unmodeled robot geometry.

Default Panda geometry

The compiled Panda retains fixed finger-joint origins at y = +/-40 mm in the hand frame. Foam supplies sphere-local centers and radii; the kinematic URDF supplies fixed-link transforms. The historical paper reference instead has +/-65 mm finger origins, shifting four finger spheres by 25 mm. These are explicitly different models, not a GRiD FK defect.

benchmark/panda_collision.py defaults to model="paper" for historical cross-solver comparisons. Use model="hjcd" for the current URDF-derived geometry. Both CPU helpers check environment collisions only; their boolean result does not certify self-collision freedom or reproduce the compiled self-pair exclusion policy. The legacy reference and production geometry are intentionally unchanged.

The native input structure and world-construction function are documented in Solver kernel.