Skip to content

CHANGELOG.md

Latest commit

 

History

History
170 lines (129 loc) · 6.75 KB

CHANGELOG.md

File metadata and controls

170 lines (129 loc) · 6.75 KB
Raw

Changelog

0.3.2 - 2026-10-18

Added

  • Astranomy Direct-API and Codex Agent-EEF policy backends, with configurable model access, action representations, history selection, and inference logs.
  • Optional episode context sent once before inference, allowing compatible policy clients to use video demonstrations for in-context learning (ICL).
  • Object-counting task predicates, multiple tracked instruction fields, and support for multi-articulation end-effector recording and gripper labels.

Changed

  • Cosmos3 supports full WebSocket URIs, optional Bearer authentication, and separate policy sessions for parallel environments. Cosmos3 and DreamZero share a MsgPack/NumPy WebSocket transport.
  • Batched policy replans are divided into bounded per-request environment groups.

Fixed

  • Ground meshes no longer occlude cleaning-scene backgrounds.
  • Recorded subtask conditionals resolve correctly during replay.

0.3.1 - 2026-08-11

Added

  • Kinova Gen3 (Robotiq 2F-85) fixed-base support: robot config, joint-position environment registrations, and an example runner (examples/run_kinova_jointpos.py). (#23)
  • Floor-standing robot support: a robot config can declare the root_z_above_ground label (the reach of its lowest colliders below the root), and the environment factory places the root on each task scene's authored ground plane plus that offset (docs/robots.md).
  • Per-scene ground-height lock (tests/test_scene_ground.py): the task table is a dynamic rigid body resting on the ground plane, so the authored ground height sets the tabletop height and is now locked per scene.

Changed

  • Every robot config must now declare the ee_recorder_bodies label mapping HDF5 channel names to the articulation bodies recorded for end-effector pose ({} opts out). Breaking for custom robot configs: environment generation fails with a ValueError naming the config until the label is added — see docs/robots.md#end-effector-pose-recording.
  • Remaining task-scene ground planes aligned to the canonical -0.697 height; legacy scenes keep their original -0.65 ground to preserve replay compatibility with existing recordings.

0.3.0 - 2026-08-07

Added

  • Explicit coordinate-frame contract: end-effector observations are now published in the robot-root frame, with a compatibility shim for older recordings (robolab/core/logging/frame_compat.py) and a reference doc (docs/frames.md).
  • Ground-truth state export during stepping: --enable-gt-state makes every runner emit a raw per-env world-state snapshot each step under obs["gt_state"] (robolab.eval.GroundTruthStateExporter; schema in docs/environment_run.md).
  • VoLo policy backend (policies/volo/): a proxy wrapper over existing clients that adds depth observations, camera calibration, and ground-truth state metadata to the inference payload.
  • Robot-owned scene fixtures with per-robot fixture selection (robolab/core/environments/scene_fixture.py): a robot config declares the table fixture it is mounted on (or none, for robots with their own base), and the environment factory swaps the task scene's default fixture accordingly.
  • Multi-gripper support in the task conditional system: robots can declare named gripper groups (e.g. "gripper": ["left", "right"] on a bimanual robot), and success/failure predicates can target any member of a group or require several grippers simultaneously (see docs/task_conditionals.md and docs/event_tracking.md).
  • Pull request template and documentation of accepted PR types in CONTRIBUTING.md.
  • News section in the README and an ecosystem page (docs/ecosystem.md).

Fixed

  • Explicitly-named tasks now resolve correctly from scoped content-pack folders.

0.2.1 - 2026-07-20

Added

  • Faithful replay of recorded episodes: new robolab/core/replay/ module with recorded env-config overlay, initial-state restore, and per-step state validation; driver script examples/run_recorded.py; user guide docs/replay.md. (#19)
  • Batched inference support in the eval client interface.
  • The pytest install-verification suite now ships in tests/ (documented fresh-install check: uv run pytest tests/).

Changed

  • Subtask progress tracking is enabled by default during evaluation.
  • Updated the GR00T N1.7 client. (#15)
  • Overrode isaacsim's stale websockets==12.0 pin so uv lock resolves a modern websockets for the policy clients.

Fixed

  • WorldState extras handling on IsaacLab 2.3 (XformPrimView).

0.2.0 - 2026-07-07

Added

  • IsaacSim 5.1 / IsaacLab 2.3 support alongside 5.0 / 2.2, selected via the mutually-exclusive isaac50 / isaac51 extras.
  • Relative differential-IK environment registrations (examples/run_rel_ik_demo.py).
  • Realtime/pathtracing RTX renderer selection via --rendering-type.
  • Poly Haven attribution files for the curated backgrounds.

Changed

  • Bundled backgrounds trimmed to a curated indoor set. More backgrounds can be obtained from Poly Haven.
  • End-effector pose position is recorded in the env-local frame.
  • Docs: asset contribution guidelines, conditionals restructure; generalized VRAM sizing guide.

Fixed

  • HDF5 episode success is derived from terminated rather than a named termination term.

0.1.1 - 2026-05-31

Added

  • Cosmos3 policy client; refreshed DreamZero, GR00T, and Pi0-family clients.
  • Results dashboard (robolab-dashboard CLI).
  • /robolab-scenegen and /robolab-taskgen Claude Code skills.
  • Convex-hull placement predicates, adaptive sampling, and DROID IK control.
  • CONTRIBUTING.md and a uv-based install flow.

Changed

  • Per-policy reorganization: one runner per backend under policies/<policy>/run.py.
  • Docs revamp: per-policy READMEs, SPDX headers, THIRD_PARTY_NOTICES.md.
  • Episode videos encoded as streaming H.264 (libx264) so they play in browsers.

Fixed

  • Placement order enforced in then-sequenced tasks.
  • Frozen-env data leaks in the episode recorder.

0.1.0 - 2026-04-10

Initial public release: 100+ benchmark tasks, DROID and Franka embodiments, server-client policy evaluation (Pi0 family, GR00T, DreamZero), and HDF5 episode recording with analysis tools.