Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

Orbit: A Unified Simulation Framework for Interactive Robot Learning Environments

Mayank Mittal, Calvin Yu, Qinxi Yu, Jingzhou Liu, Nikita Rudin, David Hoeller, Jia Lin Yuan, Ritvik Singh, Yunrong Guo, Hammad Mazhar, Ajay Mandlekar, Buck Babich, Gavriel State, Marco Hutter, Animesh Garg

Seminal January 2023

Franka arm with an Allegro hand, simulation (left) next to the real robot (right).

TL;DR An open-source framework on NVIDIA Isaac Sim for building robot learning environments with GPU-parallel rigid and deformable physics and photo-realistic rendering. It ships with 16 robots, 10 motion generators and over 20 tasks, and simulates about 10x faster than CPU-based frameworks on rigid-body tasks and 3x faster on cloth.

What they do

Orbit is an open-source robot learning framework built on NVIDIA Isaac Sim. It combines GPU-accelerated rigid and FEM-based deformable physics (PhysX 5) with ray-traced rendering and sensors such as RGB-D, semantics, LiDAR and contact. It ships with 16 robotic platforms, 10 motion generators and benchmark tasks: eleven rigid-object, thirteen deformable-object and two locomotion environments. One framework supports RL, imitation learning, motion planning and real-robot deployment.

How it works

  • World and agent. The world holds robots, sensors, objects and visualization markers, built from scripts, scanned meshes or the Isaac Sim GUI. The agent is a graph of perception and action nodes. A task adds goals, rewards and resets, so one world can serve many tasks.
  • Sim-to-real hooks. Actuator models (DC motors, series elastic actuators) inject delays and torque saturation before joint commands. Sensors and nodes each run on their own timer, e.g. IK at 50 Hz feeding a 1000 Hz joint controller.
  • Batteries included. GPU differential IK, operational-space and joint control; CPU RMP-Flow and OCS2; keyboard, gamepad and Spacemouse teleoperation; wrappers for rl-games, RSL-rl, stable-baselines3 and robomimic.
  • Deployment via ZMQ (Franka, 60 Hz commands upsampled to 1000 Hz with a quintic interpolator) or ROS (ANYmal-D).

Results

  • Throughput on an RTX 3090: about 10x faster than CPU-vectorized frameworks on rigid tasks and about 3x faster than DEDO on cloth. The CPU frameworks crash at around 200-300 environments for lack of memory.
  • PPO with RSL-rl or rl-games runs at 50,000-75,000 FPS with 2048 environments, versus 6,000-18,000 FPS with stable-baselines3.
  • Franka-LiftCube from 2000 collected demos per setting: BC-RNN succeeds 87% with randomized start and goal, versus 43% for BC.
  • An ANYmal-D locomotion policy trained only in simulation runs on the real robot. A simulated silicone beam’s damped oscillations closely follow motion-capture data.

Limitations

Physics reaches 125,000 FPS, but camera rendering tops out at 270 FPS total for ten 640x480 cameras. Simulator fidelity is checked only on one beam and qualitative sim-to-real runs. There are no tactile or force-torque sensors, and assets must be converted to USD.

Videos

Franka with its parallel gripper, commands from Orbit executed on the real robot.
Reinforcement learning workflow with many Franka environments simulated in parallel.
Full video (YouTube)

Figures

From Figure 1: The robots, sensors, rigid and deformable objects, motion generators and teleoperation interfaces incl...
From Figure 1: The robots, sensors, rigid and deformable objects, motion generators and teleoperation interfaces included in Orbit.
From Figure 2: Orbit splits a setup into a World, analogous to the real world, and an Agent, a computation graph of o...
From Figure 2: Orbit splits a setup into a World, analogous to the real world, and an Agent, a computation graph of observation- and action-processing nodes. A task is a graph cut over this graph plus a goal; for RL it defines the observation, action and reward signals.
From Figure 11: Simulation throughput on Franka-LiftCube versus number of parallel environments. CPU-vectorized frame...
From Figure 11: Simulation throughput on Franka-LiftCube versus number of parallel environments. CPU-vectorized frameworks (RoboSuite, ManiSkill2) are limited by memory and scale poorly; Orbit performs on par with IsaacGym, which uses the same physics engine.
Abstract

We present Orbit, a unified and modular framework for robot learning powered by NVIDIA Isaac Sim. It offers a modular design to easily and efficiently create robotic environments with photo-realistic scenes and high-fidelity rigid and deformable body simulation. With Orbit, we provide a suite of benchmark tasks of varying difficulty -- from single-stage cabinet opening and cloth folding to multi-stage tasks such as room reorganization. To support working with diverse observations and action spaces, we include fixed-arm and mobile manipulators with different physically-based sensors and motion generators. Orbit allows training reinforcement learning policies and collecting large demonstration datasets from hand-crafted or expert solutions in a matter of minutes by leveraging GPU-based parallelization. In summary, we offer an open-sourced framework that readily comes with 16 robotic platforms, 4 sensor modalities, 10 motion generators, more than 20 benchmark tasks, and wrappers to 4 learning libraries. With this framework, we aim to support various research areas, including representation learning, reinforcement learning, imitation learning, and task and motion planning. We hope it helps establish interdisciplinary collaborations in these communities, and its modularity makes it easily extensible for more tasks and applications in the future.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.