isaac-sim/IsaacLab

Features

• Omniverse-PhysX Backends - Provides the core Omniverse-PhysX simulation backend that models rigid bodies, articulated systems, deformable objects, and sensors.
• Multi-Agent Training Environments - Trains multiple agents simultaneously in shared environments using cooperative or competitive algorithms.
• Multi-Agent Training - Trains multiple agents simultaneously in a shared environment using cooperative or competitive algorithms.
• Reinforcement Learning Training - Trains reinforcement learning agents using the SKRL library with PyTorch or JAX backends.
• Reinforcement Learning Environments - Provides over 30 pre-built training environments for reinforcement learning, supporting single-agent and multi-agent setups.
• Multi-Agent Orchestration - Orchestrates simultaneous policy updates for multiple agents in shared environments using cooperative or competitive algorithms.
• Reinforcement Learning Training - Trains reinforcement learning agents in over 30 ready-to-use simulated environments with multi-agent support.
• RL Training Workflows - Trains reinforcement learning agents using the RL-Games library on simulation tasks.
• Library Abstraction Layers - Wraps multiple reinforcement learning libraries behind a unified training interface for drop-in swaps between RL-Games, RSL-RL, SKRL, and Stable-Baselines3.
• Multi-Agent Training - Trains multiple agents simultaneously in a shared environment using cooperative or competitive algorithms.
• Robot Policy Trainers - Trains robot policies using reinforcement learning algorithms in simulated environments with support for locomotion, manipulation, and multi-agent tasks.
• Omniverse-PhysX Integrations - Simulates rigid bodies, articulated systems, deformable objects, and sensors using NVIDIA PhysX integrated with Omniverse for real-time physics.
• Robotic Physics and Sensor Simulators - Simulates rigid bodies, articulated systems, deformable objects, and sensors for realistic robot interaction.
• Task Environment Configurations - Defines each training environment as a modular task class specifying observation spaces, action spaces, reward functions, and termination conditions.
• Trained Agent Rollout Tools - Loads a saved or pre-trained checkpoint and runs the agent in a simulation environment for evaluation.
• Motion Imitation Policies - Trains policies to imitate pre-recorded human animations using adversarial motion priors.
• Manager-Based Environment Builders - Constructs simulation environments for reinforcement learning using manager-based or direct workflow base classes.
• Policy Distillation - Compresses large teacher policies into smaller student networks through supervised learning on teacher-generated rollouts.
• Parallel Policy Rollout Tools - Loads a saved checkpoint and runs the policy across multiple parallel simulation environments to observe behavior.
• Assembly Policy Training - Trains or evaluates a policy to insert a plug into a socket across diverse assembly tasks.
• Multirotor Control Policies - Trains policies for setpoint position control of multirotor drones.
• TensorBoard Event Generators - Opens TensorBoard to inspect logged training metrics stored in the logs directory.
• Contact-Rich - Trains policies for precision assembly tasks like peg insertion and gear meshing.
• Assembly Policy Training - Generates disassembly trajectories as demonstrations and trains policies for assembly tasks.
• Locomanipulation Policies - Trains policies coordinating both locomotion and upper-body manipulation for humanoid robots.
• Locomotion Policies - Trains policies for velocity command tracking on flat or rough terrain with legged robots.
• Manipulation Policies - Trains policies for fixed-arm tasks like reaching, lifting, stacking, and opening drawers.
• Robot Model Libraries - Provides a library of over 16 robot models including manipulators, quadrupeds, and humanoids.
• Domain Randomizations - Varies scene properties like lighting, textures, and physics parameters to improve policy robustness.
• Instability Debugging Tools - Provides tools for diagnosing and fixing unstable robot simulations using the PhysX Visual Debugger.
• Simulation Debuggers - Diagnoses and resolves simulation issues like memory leaks, GPU buffer overflows, physics instability, and NCCL communication errors.
• Physical Environment Simulation - Tunes simulation parameters and uses the PhysX Visual Debugger to diagnose and fix instability in robot environments.
• Robot-Attached Sensor Simulations - Attaches camera and ray-caster sensors to a robot to collect environmental observations during simulation.
• Modular Component Swappers - Swaps or extends robot models, sensor configurations, and environment logic through a modular component system.
• Robot Model Libraries - Offers a library of more than 16 robot models including manipulators, quadrupeds, and humanoids for simulation.
• Gymnasium Space Configurators - Configures environments with various Gymnasium space types like Box, Discrete, Dict, and Tuple for observation and action definitions.
• Task-Space Controllers - Commands robots at the task level using operational space or task-space controllers.
• Training Metrics - Writes structured training metrics to disk in a hierarchical directory format for real-time visualization and post-hoc analysis with TensorBoard.
• Training Metric Loggers - Writes structured training metrics to disk in a hierarchical directory format for real-time visualization and post-hoc analysis with TensorBoard.
• Reinforcement Learning - GPU-accelerated framework for robot learning.
• Robotics Simulation - Modular framework for robot learning using physics simulation.
• Robotics Simulators - Provides a high-fidelity simulation environment for robotic research.

Star history