cuRoboV2-Fast: Fast Torque-Limited Collision-Free Motion Planning

Balakumar Sundaralingam, Adithyavairavan Murali, Stan Birchfield

November 2026

Real YAM arm clearing a table at 1x, with the voxel map, stereo view and live joint torques and velocities overlaid.

TL;DR A GPU motion planner that optimizes B-spline trajectories with the payload in the dynamics, then retimes them under velocity, acceleration, jerk and torque limits. It plans a 3 kg-payload motion in 4.4 ms on an RTX 5090, 14.7x faster than cuRoboV2, and cuts real pick-and-drop cycle time by 34.6%.

What they do

cuRoboV2-Fast is a single-GPU planner for collision-free, torque-limited arm motions that cuts both planning latency and execution time. The payload is part of the dynamics, so a heavy object can change the path itself, not just slow it down.

How it works

B-spline trajectory optimization. Each trajectory sample depends on four controls, so GPU threads evaluate samples independently and each control gathers gradients only from its local support. This matches cuRoboV1’s success with 25 instead of 125 initial iterations.
TimeOpt. With the path fixed, a second cubic B-spline over squared path speed is optimized under sampled velocity, acceleration, jerk and torque limits. Local support gives banded Newton systems, solved by a primal-dual predictor-corrector method in 0.28 ms on average.
Faster robot computations. Shared gradient accumulation for kinematics, per-link bounding boxes that skip sphere-pair collision tests, and differentiable inverse dynamics whose payload can change at runtime without recompiling.
Parallel GPU RRT seeds the optimizer when a straight line fails, in 0.305 ms on average.

Results

With a 3 kg payload, planning takes 4.43 ms on an RTX 5090 (14.7x faster than cuRoboV2) and 25.99 ms on a Jetson AGX Orin, faster than cuRoboV1 + TOPPRA on the RTX 5090. Success matches cuRoboV2 and beats the TOPPRA baselines by 14.5-15.7 points at the heaviest payload.
TimeOpt shortens mean motion duration by 15.7% beyond uniform timestep refinement.
On a real YAM arm doing table clearing and bin picking from stereo cameras, mean cycle time is 5.93 s, 34.6% shorter than cuRoboV2; planning takes 1.0% of it.

Limitations

TimeOpt converges to a local optimum, with no guarantee of time-optimality. With planning down to 1% of the cycle, the authors name reliable grasping and contact-rich interaction as the remaining bottleneck.

Videos

Picking objects out of a cluttered bin on the real robot.

Reacting to scene changes as a person pushes a pole into the workspace during a transfer.

Figures

From Figure 17: The real-robot system on one RTX 6000 Ada GPU. SceneFields builds a TSDF for GraspGenX grasp generati... — From Figure 17: The real-robot system on one RTX 6000 Ada GPU. SceneFields builds a TSDF for GraspGenX grasp generation and an ESDF for collision checking; cuRoboV2-Fast selects a grasp and plans the pregrasp, approach and lift motions in 23 ms, with RRT initialization when needed.

From Figure 4: TimeOpt keeps the position spline fixed and optimizes a second spline for squared path speed, so equal... — From Figure 4: TimeOpt keeps the position spline fixed and optimizes a second spline for squared path speed, so equal path intervals get unequal traversal times: slower through the shaded region and faster elsewhere, for a shorter total duration.

From Figure 18: Arm utilization, motion time and cycle time for real-robot object transfers. cuRoboV2-Fast nearly dou... — From Figure 18: Arm utilization, motion time and cycle time for real-robot object transfers. cuRoboV2-Fast nearly doubles the time spent at 90-100% velocity utilization and cuts motion and cycle times by 44.8% and 34.6% relative to cuRoboV2.

Abstract

We present cuRoboV2-Fast, a GPU-accelerated planner that reduces both planning latency and execution time for collision-free, payload-aware, torque-limited motions. B-spline trajectory optimization reduces the required iterations, while accelerated robot computations and differentiable inverse dynamics reduce refinement cost and accommodate changing payloads. TimeOpt optimizes traversal speed along the fixed path under sampled velocity, acceleration, jerk, and torque constraints, exploiting local spline support to solve banded Newton systems efficiently on the GPU. With a 3 kg payload, cuRoboV2-Fast achieves 4.4 ms mean planning time on RTX 5090, 14.7x faster than cuRoboV2. Our method on Jetson AGX Orin plans faster than cuRoboV1 + TOPPRA on RTX 5090. A stereo-camera-based deployment on YAM achieves 5.93 s mean completed pick-and-drop cycles, 34.6% shorter than cuRoboV2. Code (Apache 2.0): www.github.com/NVlabs/curobo , Project Page: www.research.nvidia.com/labs/lpr/curobov2-fast .

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.

Matt Strong