A new preprint describes a GPU simulator that, in pure stepping, reached 2.03 million environment steps per second in a simple CartPole test with 8,192 parallel environments. In a SoftBodyPush test, the reported rate was 8.86 thousand environment steps per second and leveled off at around 512 environments. The contrast shows how sharply the workload can change when a benchmark moves from a rigid-body task to a deformable one.
The results come from CRESSim-Neo, a batched GPU engine that combines position-based simulation of rigid bodies, deformable tissues, fluids and strands with batched rendering and surgery-specific sensing. Its data path is GPU-resident, with DLPack-based zero-copy access to PyTorch, a way to expose buffers without an extra copy between those components.
One engine, several kinds of scenes
Under the hood, CRESSim-Neo is a standalone C++ engine with low-level GPU execution and high-level Python bindings. For surgical scenarios, the system includes routed cable constraints, tool, needle and thread attachments, and path-following suturing constraints.
To test its range, the authors ran seven named simulated reinforcement-learning environments on a single GPU. Four were simple tasks: CartPole, SoftBodyPush, FluidPour and TargetCenter. Three were surgical scenes: TissueRetract, BloodSuction and UltrasoundScan. Throughput was reported in two modes, pure stepping across parallel environments and end-to-end PPO, including rollout collection and policy optimization.
The speed depends on the workload
CartPole delivered the headline pure-stepping result: 2.03 million environment steps per second with 8,192 parallel copies. The task uses the engine's rigid-body simulation.
SoftBodyPush showed a different scaling pattern. It reached 8.86 thousand environment steps per second and then saturated at around 512 environments.
On most of the heavier particle-based and surgical tasks, end-to-end PPO throughput stayed close to the pure-stepping rate. The comparison was descriptive, however. The paper reports no confidence intervals, replicate counts or variability estimates for the throughput results.
Surgical demonstrations, not surgical proof
Beyond the speed tests, the engine demonstrated simulated workflows for tissue retraction, fluid suction and ultrasound targeting in deformable scenes. These examples show software support for those task types, but they are demonstrations of simulated scenes and sensor outputs, not tests of clinical performance.
The authors put the reinforcement-learning examples in a narrower category. They were intended to demonstrate the learning interface and computational performance, not the quality of a learned policy. Task-specific learning and transfer were outside the study's scope.
The boundaries of the result
The simulator uses position-based dynamics, or PBD, a simulation method built around object positions. The authors state that PBD is less suitable for applications requiring accurate rigid-body contact dynamics.
Scaling also brings memory and scene-management constraints. Buffers are allocated according to per-environment capacity, while topology-changing operations require buffer reallocation, scene rebuilding or re-uploading the simulation state.
The reported speed is tied to a test setup using an Intel Core i9-14900K CPU, an NVIDIA GeForce RTX 4090 GPU and 29 GiB of host memory. Cross-platform validation remains incomplete. The active backends are Vulkan and D3D12, ultrasound currently requires CUDA interoperability, and the renderer lacks ray tracing and custom graphics pipelines or shaders.
Taken together, the work presents a unified software platform and reports high throughput in selected simulated workloads, but leaves the central real-world questions open. The evidence does not establish clinical effectiveness, improved surgical outcomes, learned-policy performance or sim-to-real transfer. Nor does it establish superiority over existing simulators through a head-to-head comparison.
The manuscript states that it was submitted to IEEE for possible publication. The research reports support from the Canada Foundation for Innovation, NSERC Canada, the Canadian Institutes of Health Research and the Government of Alberta Major Innovation Fund A-MEDICO.
Paper data and sources
Original title: CRESSim-Neo: A Batched GPU Simulation Engine for Surgical Robotics and Robot Learning
Authors: Yafei Ou, Ahnaf Naheen, Tleukhan Mussin et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text