Release v0.6.0

[sebcrozet]

v0.6.0 (4 October 2026)

Breaking changes

• ⚠ Per-body, per-collider and joint GPU buffers are batch-interleaved: entity i of env e is at
  i * num_batches + e (was e * stride + i). This affects RbdState::body_poses and the other
  per-body tensors, and GpuRigidBodyRef::gpu_id.
• ⚠ Collision pairs and contacts are one flat buffer shared by all envs: collision_pairs_len and
  NexusCounts::collision_pairs report the total, RbdCapacities::collisions_capacity is a
  per-env hint, and RbdState::contacts_len is removed.
• ⚠ RbdSimParams is a single uniform shared by all envs (finalizing panics if envs differ). New
  fields: contact_merge_cos, num_internal_pgs_iterations, friction_in_bias_pass.
• ⚠ The contact prediction distance is RbdSimParams::prediction_distance() instead of a
  hard-coded 0.02.
• ⚠ GpuMultibodySet::set_num_internal_pgs_iterations no longer drives the solver; use
  NexusState::set_rbd_num_internal_pgs_iterations (or RbdSimParams::num_internal_pgs_iterations).
• ⚠ GpuMultibodySet: dof_values, contact_constraints_per_batch, contact_constraint_jacs and
  contact_constraint_columns are removed (see contact_jac_cols); from_rapier takes a
  contact-slot count.
• ⚠ Viewer: UiSections is replaced by the UiSection enum (UiState::ui_section); one panel tab
  is open at a time.
• ⚠ MPM shaders: NBH_SHIFTS is now a plain array; use nbh_shift(i) for a vector.
• Many low-level kernel argument structs and dispatch signatures changed (SolverArgs,
  ColoringArgs, GpuNarrowPhase::dispatch, BatchIndices, ...).
• Built against khal 0.4, vortx 0.5, kiss3d 0.47, rapier 0.36 and parry 0.31.

Added

• NexusState::set_deterministic: two runs of the same scene give identical results on the same
  machine and build. Costs ~10-17% per step, up to ~45% with trimeshes or polylines, and fixes the
  rigid-body resize policies. Testbed: "Deterministic" checkbox, --deterministic flag and a step
  counter (NexusState::steps).
• cuda-oxide feature (CUDA kernels compiled by cuda-oxide), and compute graphs:
  NexusState::set_compute_graphs_enabled records the step once and replays it as a CUDA graph
  (testbed: --compute-graphs). Both need the GitHub version of khal/vortx.
• Viewer debug renderer ("Debug render" tab): collider shapes, AABBs, joints, contacts, LBVH
  nodes, MPM particles and grid. Readbacks: RbdState::debug_contacts, debug_lbvh,
  MpmState::debug_particles, debug_grid.
• Viewer sensor cameras (3D): add_sensor_camera, attach_sensor_camera, render_sensor_rgb,
  render_sensor_depth, render_sensor_segmentation, with MSAA, shadow and ambient controls.
• Viewer: set_ambient, add_directional_light, set_body_color, set_body_casts_shadows,
  textured visual meshes, set_shadow_softness/range/resolution.
• Env reset (3D): RbdState::snapshot, reset_env_from_snapshot, and the batched
  publish_reset_templates/reset_envs_from_templates.
• Multibodies: joint dry friction (MJCF frictionloss), per-link contact sensors
  (set_contact_sensor_links), actuator delay, motor targets read from a GPU tensor
  (scatter_motor_targets), and set_substep_refresh.
• NexusState: control_multibody_motors, read_multibody_links,
  multibody_joint_positions/set_multibody_joint_positions/multibody_joint_velocities,
  read_rigid_body_poses/velocities and set_rigid_body_pose/velocity (3D),
  set_rbd_timestep, set_rbd_collisions_capacity, set_rbd_mb_contact_constraints_capacity,
  set_rbd_implicit_coriolis, set_rbd_substep_refresh.
• RbdSimParams::friction_in_bias_pass: also solve friction during the biased pass (off by
  default), which keeps pinch grasps and resting stacks from drifting.
• Optional contact reduction to at most 4 points per collider pair (RbdPipeline::contact_reduction).
• RbdPipeline::step_encoded records a step into a caller-owned encoder.
• Python: Robot (URDF/MJCF loading, robot_state/robot_qvel, set_robot_targets, FK/IK, PD
  gains), MJCF actuators (apply_actuator_controls), body pose/velocity read/write,
  set_rbd_solver_params, sensor cameras, contact debug readbacks, set_compute_graphs_enabled.

Fixed

• Bodies resting on trimeshes were ejected (unvalidated seeded colors, EPA picking a triangle back
  face, duplicate manifold points).
• Warmstarting matched contacts by body pair and to the first old point within 10 cm; it now
  matches by collider pair and sub-shape, to the nearest point. Also fixes polyline warmstart.
• Contacts between a multibody link and a rigid body were solved twice, cancelling a robot's grip.
• Multibody contacts beyond 128 points per multibody were silently dropped.
• The narrow phase used env 0's collider parents in every env.
• Polyline AABBs ignored the segment thickness.
• GpuMultibodySet::read_dof_coords returned wrong locked axes for envs other than 0.
• Buffer auto-resize could race in-flight steps on Metal and leave stale warmstart ranges.
• The crates did not build without the mpm feature.
• RbdState::debug_constraint_colors now reads the constraints and colors of the last step.

Modified

• Performance: batched kernels run as flat dispatches over the interleaved layout, pair buffers
  are sized by total demand instead of the busiest env, and multibody contacts are indexed per
  multibody.
• Multibody contact friction is velocity-level only, and CFM softening applies only to normal rows.
• num_internal_pgs_iterations also drives the rigid-body sweeps of the biased pass, interleaved
  with the multibody sweeps.
• Viewer defaults: 4096 shadow map over 4 atlas layers, first cascade at 3 m, mipmapped textures
  with 16x anisotropic filtering. The viewer no longer steps before the first sync of a scene.