ORBIT-Surgical: An Open-Simulation Framework for Learning Surgical Augmented Dexterity

*Equal Contribution
1University of Toronto, 2University of California, Berkeley, 3ETH Zurich,
4Georgia Institute of Technology, 5NVIDIA
IEEE International Conference on Robotics and Automation (ICRA) 2024


Physics-based simulations have accelerated progress in robot learning for driving, manipulation, and locomotion. Yet, a fast, accurate, and robust surgical simulation environment remains a challenge. In this paper, we present ORBIT-Surgical, a physics-based surgical robot simulation framework with photorealistic rendering in NVIDIA Omniverse. We provide 14 benchmark surgical tasks for the da Vinci Research Kit (dVRK) and Smart Tissue Autonomous Robot (STAR) which represent common subtasks in surgical training. ORBIT-Surgical leverages GPU parallelization to train reinforcement learning and imitation learning algorithms to facilitate study of robot learning to augment human surgical skills. ORBIT-Surgical also facilitates realistic synthetic data generation for active perception tasks. We demonstrate ORBIT-Surgical sim-to-real transfer of learned policies onto a physical dVRK robot.

Short Video

Long Video

Reinforcement Learning

ORBIT-Surgical supports various RL frameworks such as RSL-RL and RL-Games to train RL policies for surgical tasks involving rigid and soft objects.

RL Policy for Shunt Insertion

Example Tasks

Supplementary Material

S1: Multi-stage Imitation Learning

Long-horizon tasks are computationally expensive and can be hard to be solved by plain Imitation Learning (IL) algorithms. Instead, multi-stage IL policy can be learned efficiently, where long-horizon tasks are divided into subtasks, and an IL policies are trained to perform each subtask separataely. All subtasks are then chained for end-to-end policy execution. Here, we divide a "Needle Pick and Transfer" task into three stages, i.e. Pick, Handover, and Reach, and trained subpolicies to perform each subtask using behavior cloning. Figure below shows a successful end-to-end policy execution in simulation.

Multi-stage imitation learning.

DualArm - Needle Pick and Transfer task with multi-stage imitation learning. a) starting state of both PSMs, b) right PSM moves down to pick up the needle, c) needle handoff occurs between the PSMs, d) PSMs move away from each other, e) needle is fully transferred to other arm.

S2: System and Physics Parameters

We list, categorize and conduct a sensitivity analysis on the simulation parameters in the NVIDIA Isaac Sim. We focus on the representative environment, 'Threaded Needle Pass Ring', where the task is to pass a suture needle with a soft thread through a ring. This environment is selected because it involves extensive interactions between objects, including interactions between deformable objects and rigid objects. To determine the impact of individual parameters, we held all others constant, varied one specific parameter at a time, and noted results.

System parameters:

  • dt (The physics simulation time-step in seconds). We currently use the value of 0.005 because this is the largest dt value that can provide a fast and stable simulation without jitter. We find that dt of 0.007 cause oscillation when objects interact and deformable object simulation is not stable. We find that the smaller dt value of 0.002 results in a stable yet slower simulation, extending the time required for a single demonstration. For example, the time needed to collect one demonstration is 27 seconds for dt 0.005 and 68 seconds for dt 0.002.
  • Solver type [PGS (Projective Gauss-Seidel) or TGS (Truncated Gauss-Seidel)]. Both PGS and TGS offer similar levels of stability and realism in simulations. We used TGS.

Physics parameters - Rigid body (suture needle):

  • Solver iteration counts: We find this parameter is not a sensitive parameter. Solver iteration counts with value 1 and a much larger value, 128 can provide similar simulation behavior. But a large solver iterations count slows the simulation process. We currently use 1 for both solver_position_iteration_count and solver_velocity_iteration_count, which is the minimum value that ensures stability.
  • Contact offset and Rest offset: They are used to control the contact generation. Collision detection starts to generate contact points as soon as two shapes get closer than the sum of their contact offsets. The rest offset quantifies how close a shape gets to others at rest. An “auto compute” feature is provided by Nvidia Isaac Sim. By default, the two offsets, remeshing resolution and triangle count can be all computed automatically by the engine based on the objects’ mesh geometry. These auto computed values can sometimes provide a stable simulation, sometimes will just give bad simulation, depending on the assets’ geometry. We find that the autocomputed contact offset and rest offset already provide a stable and realistic for our suture needle simulation.
  • Friction: (both static and dynamic friction). We currently use the value of 0.5 for both static and dynamic friction.
  • Damping: We currently have linear damping of 0.05 and angular damping of 0.05. We find the damping does not have much impact on the stability of the simulation. But larger damping slows down the movement of the rigid body. We use the minimum value that ensures stability.

Physics parameters - Soft body (thread):

  • Simulation mesh resolution: The simulation mesh resolution is highly sensitive. We currently use the value 28 for the thread. Increasing or decreasing the parameter by a value greater than 3 can lead to a badly simulated collision mesh or mismatch between visual mesh and collision mesh. The simulation mesh resolution needs to be carefully tuned.
  • Contact offset and Rest offset: contact offset and rest offset are used to compute the contact with other objects. For the deformable thread, the autocomputed offsets lead to unstable simulation and the offset parameters of the thread are very sensitive. We currently use 0.0003 for contact offset and 0.0002 for rest offset. Increasing or decreasing 0.0001 can lead to mismatch between visual mesh and collision mesh, which can lead to the issue of penetrating other objects during interactions.
  • Solver position iteration count: We find this parameter is not a sensitive parameter. Solver position iteration counts with value 1 and a much larger value, 128 can provide similar simulation behavior. But a large solver iterations count slows down the simulation process. We currently use 16 for solver_position_iteration_count, which is the minimum value that ensures stability.
  • Remeshing resolution and triangle count: These parameters are used for computing the collision mesh by Nvidia Isaac Sim. We use the auto-computed remeshing resolution and triangle count which gives a stable simulation.

Frequently Asked Questions

Q: Is ORBIT-Surgical fully free and open-source?

A: The underlying robotics simulation application, NVIDIA Isaac Sim is free for individual use. ORBIT-Surgical will be released as a free and open-source package upon publication.

Q: What are the differences between ORBIT-Surgical and Intuitive Surgical's proprietary SimNow?

A: SimNow is a commercial platform for surgeons to practice the skills and techniques needed to perform surgical procedures. It provides a simulated environment that can give the surgeon a realistic experience; ORBIT-Surgical is intended as a benchmark to spur research on supervised autonomy for robot-assisted surgery. We aim to provide simulated environments with a large number of parallel environments for efficient data collection and fast surgical robot learning, as well as accurate physics for sim-to-real transfer of learned policies. In addition, the differences also lie in the way how the simulators can be utilized. Specifically:

  • While SimNow is tailored for human practitioners using the MTMs, ORBIT-Surgical offers integration with a range of reinforcement frameworks for learning.
  • SimNow is compatible with da Vinci Si, X/Xi, and SP. However, it does not support the da Vinci Research Kit (dVRK) which is commonly found in academic labs.
  • SimNow is not open-source while ORBIT-Surgical will be open-sourced upon publication to contribute to the community.

Q: How much time is required to train the RL policies?

A: Both tasks involving "Suture Needle Lift" task and "Shunt Insertion" task can be trained under 2 hours on a single NVIDIA RTX 3090 GPU.

Our Related Projects

SuFIA: Language-Guided Augmented Dexterity for Robotic Surgical Assistants
Masoud Moghani, Lars Doorenbos, William Chung-Ho Panitch, Sean Huver,
Mahdi Azizian, Ken Goldberg, Animesh Garg
website / video

We present SuFIA, a framework for natural language-guided augmented dexterity for robotic surgical assistants. SuFIA incorporates the strong reasoning capabilities of large language models (LLMs) with perception modules to implement high-level planning and low-level control of a robot for surgical sub-task execution.


For any questions, please feel free to contact Masoud Moghani and Animesh Garg.