Skip to content

Releases: nvidia-isaac/WBC-AGILE

v1.2: G1 Motion Tracking, Pick-and-Place, and Sim2MuJoCo

24 Mar 19:07
@huihuaNvidia2023 huihuaNvidia2023
Immutable release. Only release title and notes can be modified.
v1.2
ad4ff6e
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Choose a tag to compare

View all tags
v1.2: G1 Motion Tracking, Pick-and-Place, and Sim2MuJoCo Latest
Latest

AGILE v1.2 Release Notes

Highlights

  • G1 Motion Tracking Task — Full-body motion tracking for Unitree G1 using reference trajectories, with support for AMASS retargeted motion data
  • G1 Pick-and-Place Task — Loco-manipulation task combining trajectory tracking with object grasping via the Dex3-1 hand
  • Sim2MuJoCo — Unified sim-to-sim transfer workflow supporting both velocity tracking and motion tracking policies in MuJoCo, with a built-in evaluation pipeline for automated sweeps
  • GR00T Data Recording & Inference — Record demonstrations in Isaac Lab for GR00T fine-tuning and deploy with NVIDIA GR00T
  • OSMO Cloud Training — Workflow pipeline for launching distributed training on NVIDIA OSMO
  • Sphinx Documentation — Full documentation site with GitHub Pages deployment

New Features

Tasks

  • G1 Motion Tracking (G1-MotionTracking-v0): Whole-body motion imitation using reference trajectory tracking with body position, orientation, and joint-level rewards (#41)
  • G1 Pick-and-Place (G1-PickPlace-Tracking-v0): Trajectory-guided pick-and-place with the Unitree G1 and Dex3-1 hands, including object interaction rewards, curriculum learning, and domain randomization for data recording
  • T1 Stand-Up (T1-StandUp-v0): Get-up-from-ground task for the Booster T1 with adaptive lift curriculum (#25)

Algorithms & Training

  • Good/bad termination handling with value bootstrapping for improved PPO training (#40)
  • Reward normalization for more stable training across reward scales
  • OSMO workflow training pipeline for cloud-based distributed training (#33)

Sim2MuJoCo

  • Unified sim2mujoco module supporting both velocity tracking and motion tracking policies (#42, #44)
  • Built-in evaluation pipeline with configurable command schedules, automated sweeps (velocity, height, yaw rate), and data logging

GR00T Integration

  • Data recording pipeline: record demonstrations from trained policies with visual domain randomization
  • convert_to_gr00t.py for converting recorded data to GR00T-compatible format
  • GR00T inference service for deploying models back into the environment (#28)

Debugging & Visualization

  • Object pose GUI and reward visualizer for interactive debugging of manipulation tasks
  • Debug GUI gain fix for joint position control (#25)

Documentation & CI

  • Sphinx documentation site with NVIDIA theme, deployed to GitHub Pages (#46)
  • GitHub Actions CI with pre-merge checks and docs deployment
  • Pull request template (#34)

Bug Fixes

  • Fix pick-place action scale — Upper-body action scale filter was comparing joint names against regex patterns using exact string matching, causing right arm joints to be frozen (scale=0) during training (#47)
  • Fix velocity-height G1 training configuration (#29)
  • Fix reward normalization in distillation script
  • Fix debug GUI play mode to not require a policy checkpoint
  • Fix CI timeout for stand-up task (#38)
  • Misc workflow and configuration fixes (#45)

Infrastructure

  • Update Isaac Lab Docker image from 2.3.0 to 2.3.1
  • Add scripts/utils/convert_retargeted_data_for_tracking.py for converting AMASS retargeted motion data to the tracking task format
  • Add evaluation framework with configurable command schedules, trajectory logging, and W&B integration
Assets 3
Loading

v1.1: OSMO Workflow, GR00T Pipeline & Object Interaction

05 Mar 18:26
@huihuaNvidia2023 huihuaNvidia2023
Immutable release. Only release title and notes can be modified.
v1.1
166ee36
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Choose a tag to compare

View all tags
v1.1: OSMO Workflow, GR00T Pipeline & Object Interaction

Release v1.1

This release brings several major features including the new OSMO workflow training pipeline, the GR00T data recording pipeline, and a robot object interaction task. It also includes various bug fixes, visualizer tools for debugging, and significant CI/CD enhancements.

🚀 Features & Enhancements

  • OSMO Integration: Added OSMO workflow training pipeline (#33).
  • GR00T Data Pipeline: Added GR00T data recording, conversion, and inference pipeline (#28).
  • Robot Object Interaction: Introduced a new robot object interaction task.
  • Sim2Sim Module: Added the sim2sim module for sim-to-sim validation.
  • Visualizers: Added object and reward visualizers for easier debugging.
  • Reward Normalization: Implemented reward normalization features.
  • T1 Stand-up Task: Fixed and enhanced the T1 stand-up task (#25).
  • Recurrent Student Policy: Uploaded the PyTorch script for the recurrent student policy.

🐛 Bug Fixes

  • Training Fixes: Fixed the velocity height for G1 training (#29).
  • Debug GUI: Fixed an issue where the gain was not effective in the debug GUI, and resolved general GUI play issues.
  • Debug Env: Fixed the debug environment and updated the play script to not require a policy.
  • Distillation Script: Addressed an issue with missing reward normalization in the distillation script.

🛠️ Maintenance & CI/Misc

  • CI/CD Improvements:
    • Increased timeout for the standup task in CI (#38).
    • Updated the Pull Request template to reflect enabled CI/CD (#34).
    • Enabled CI and fixed the sim2mujoco test (#30).
  • Docker Image: Updated the Isaac Lab Docker image from 2.3.0 to 2.3.1.
  • Documentation: Added office hour FAQs and related links.
Loading