Skip to content

CHANGELOG.md

Latest commit

 

History

History
71 lines (52 loc) · 4.01 KB

CHANGELOG.md

File metadata and controls

71 lines (52 loc) · 4.01 KB
Raw

Changelog

[0.0.5] - 2026-05-19

Added

  • pick_block environment with a Holding goal for pick-only planning (#18)
  • tests/test_pick_only_goal.py covering planning with a Holding-only goal (#18)
  • tests/test_task_planning.py covering BFS goal-literal validation (#14)

Fixed

  • BFS now rejects goal atoms whose non-fabricable literals (e.g. movable, surface) do not appear in the initial state, raising ValueError instead of expanding the frontier forever. Fixes a hang triggered by hallucinated perception objects in the goal (#14)
  • Pick-only goals no longer crash in cost-function setup: object pairs and movable-to-world collision masks now handle objects that are picked but never placed (#18)

Changed

  • Consolidated fresh-symbol minting in get_valid_ground_operators behind a single _FABRICABLE_TYPE_PREFIXES table so the sampler and the BFS goal validator cannot drift out of sync (#14)

[0.0.4] - 2026-04-14

All changes in this release are from #11 (Warp sphere overlap + FK Pose optimizations).

Added

  • NVIDIA Warp kernel for sphere_to_sphere_overlap with fused cost + analytical gradients, replacing the PyTorch pairwise implementation
  • Concatenated robot_to_movables kernel launch — single call over all movable spheres instead of per-object launches
  • torch.profiler.record_function annotations through the optimization loop, cost function, and rollout
  • --torch-profile, --torch-profile-output, --coll_n_spheres, --placement_shrink_dist, and --prop_satisfying_break CLI flags on cutamp-demo
  • blocks_5 environment for benchmarking
  • tests/test_sphere_overlap.py correctness + gradient tests for the Warp kernel
  • docs/profiling-analysis.md covering how to profile cuTAMP and where time goes today

Changed

  • Rollout stores ee_position and ee_quaternion directly from cuRobo FK; kinematic_costs no longer round-trips through Pose.from_matrix
  • Removed per-step torch.cuda.synchronize() in the optimization loop that was forcing CPU-GPU pipeline stalls

Performance

End-to-end optimization loop wall time, 100 steps, 512 particles, RTX 3090, median of 3 runs. Speedup scales with the size of the movable-sphere pairwise tensor (more objects × more spheres/object → larger win):

Env Before (0.0.3) After (0.0.4) Speedup
tetris_3 (3 blocks, ~6 sph/obj) 1.46s 1.36s 1.07x
blocks (4 blocks, ~50 sph/obj) 3.43s 1.68s 2.04x
blocks_5 (5 blocks, 50 sph/obj) 4.98s 1.87s 2.66x

[0.0.3] - 2026-04-13

Added

  • max_motion_refine_attempts config option to cap the number of satisfying particles tried during motion refinement per skeleton (#9)
  • Retry next plan skeleton when motion refinement fails instead of breaking out of the loop (#9)
  • Test environment and pytest for movable-to-world initial collision fix (#10)

Fixed

  • Movable-to-world collision now masks initial timesteps per object, matching the movable-to-movable approach, so objects with perception noise at their initial pose are not incorrectly penalized (#10)
  • Vectorized movable-to-world collision into a single batched collision_fn call with cached mask for better performance (#10)
  • break_on_satisfying no longer exits the skeleton loop when motion planning is enabled but all particles fail (#9)

[0.0.2] - 2026-04-01

Added

  • Expose __version__ in __init__.py via importlib.metadata

Fixed

  • Disable CPU pose update to avoid mutating object pose in TAMPEnvironment

[0.0.1] - 2026-03-22

Added

  • Initial release of cuTAMP from NVLabs — GPU-parallelized TAMP solver with core algorithm, cost functions, rollout, samplers, task planning search, and environment definitions (book shelf, stick button, tetris)
  • Robot support for Franka Panda, FR3, and UR5e with Robotiq 2F-85/140 grippers
  • TiPToP integration: extended algorithm and motion solver, added Franka+Robotiq robot config, OBB collision utilities, new environment assets
  • Return failure reason from planner for better diagnostics
  • Log git status in experiment logger