Releases: NVlabs/COMPASS
Releases · NVlabs/COMPASS
v1.6.0
Added
- OSMO cloud submission:
osmo/run_osmo.py+ four workflow YAMLs (train / eval / record / distillation). Assets (X-Mobility base checkpoint, COMPASS USDs) download from HuggingFace inside the workflow; W&B / HF tokens routed via env vars or interactive--prompt. Includes--embodiment/--environmentoverrides for multi-embodiment training sweeps without YAML hand-edits. - Multi-GPU residual-RL training with
--distributed: torchrun-launched, per-rank Isaac Sim instance, manual gradient / KL / metric all-reduce inPPO.update, rank-0-only logger / checkpoint / video / episode-log writes. Pairs withosmo/run_osmo.py --num-gpus Nand the parameterizedosmo/workflows/rl_es_train_workflow.yaml(handles 1-N GPUs via a single{{num_gpus}}template variable). - Auto-generated occupancy maps from USD:
scripts/generate_omap_from_usd.pywraps Isaac Sim'sisaacsim.asset.gen.omap.OccupancyMapCollisionCheckerfalls back to discovering<usd_dir>/omap/occupancy_map.yamlwhenOMAP_PATHShas no entry, so SAGE-driven training and new scenes ramp up without hand-tuned omaps. - Docker-as-venv dev environment:
docker/run.sh(build / assets / up / down / exec / shell / status subcommands) plus a sourceabledocker/activatethat transparently routespython/pip/tensorboardthrough the container. Cuts first-time setup from ~30-60 minutes to ~3 minutes. - COMPASS docs site: Sphinx handbook (
docs/handbook/) deployed atnvlabs.github.io/COMPASS/docs/; academic project page consolidated underdocs/project_page/. Auto-deployed via.github/workflows/docs.yml. - Agentic skills for Claude Code under
.claude/skills/:/compass(umbrella front-door for training / eval / SAGE / OSMO / deployment) plus specialty siblings/compass-doctor(8-check diagnostic withscripts/compass_status.sh) and/compass-newembodiment(interactive robot-onboarding flow). - CI:
.github/workflows/pre-commit.ymlruns yapf / pylint / nbstripout / clang-format / large-files / trailing-whitespace / EOF / requirements-txt-fixer on every PR. - No-regression benchmark gate:
osmo/run_benchmark.pyfires one OSMO eval workflow per scene for a given embodiment (default sweep across 5 scenes); results land in W&B for offline regression assessment.
Changed
- Isaac Lab 2.1 → 3.0.0-beta1. Breaking for downstream consumers of
mobility_esAPIs:AdditiveUniformNoiseCfg→UniformNoiseCfg; restructured physics viaisaaclab_physx.physics.PhysxCfg;rerender_on_reset→num_rerenders_on_reset;ActionTermimport moves fromisaaclab.envs.mdp.actionstoisaaclab.managers; asset data fields (root_pos_w,root_quat_w, etc.) requirewp.to_torch()wrapping;write_root_state_to_sim()split intowrite_root_pose_to_sim_index()+write_root_velocity_to_sim_index(); quaternion convention flipped wxyz → xyzw acrossenvironments.pyandNonHolonomicPerfectControlActionyaw extraction;velocity_limit→velocity_limit_simon carter caster joints;commands.UniformPose2dCommandused directly (module reorg). --wandb-project-nameis now required inrun.py,record.py, anddistillation_train.py(previously defaulted to internal project names).requirements.txt: 17 previously-unpinned dependencies pinned to versions verified insidecompass-rl:latest.README.mdleads with the Docker quick-start; bare-metal install moved under "Manual install".- Pre-commit
yapfbumped to v0.43.0 for Python 3.12 compatibility (the previous v0.31.0 pin importslib2to3, removed from the stdlib in 3.13).
v1.5.2
v1.5.1
v1.5.0
[1.5.0] - 2025-07-29
Added
- Enabled goal heading alignment for object navigation integration, improving navigation accuracy to targets.
- Added ROS2 deployment examples for COMPASS Navigator, including containerized workflows and Isaac Sim integration.
- Added target_pose_generator node to interface with object localization modules to enable object navigation using COMPASS
Changed
- Upgraded Isaac Lab to version 2.1 for improved simulation capabilities and compatibility with the latest Isaac Sim and ROS2 releases.
Fixed
- Disable weights only for generalist policy loading