Fewer Iterations
B-spline trajectory optimization refines geometry with fewer iterations.
cuRoboV2-Fast: Fast Torque-Limited Collision-Free Motion Planning
- 4.4 msMean planning time on RTX 5090
- 14.7×faster than cuRoboV2
- 10.0×faster than cuRoboV1 + TOPPRA
- 34.6%shorter real-robot cycle time
Planning benchmarks at 3 kg on an RTX 5090 GPU; real-robot results on an RTX 6000 Ada.
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, 14.7× 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.
Real-robot deployment on an RTX 6000 Ada GPU.
Approach
We present cuRoboV2-Fast, a single-GPU planner for collision-free, torque-limited motions that reduces both planning latency and execution time. B-spline trajectory optimization refines geometry with fewer iterations, while TimeOpt optimizes traversal speed along the resulting fixed path. The timing spline's local support yields banded systems for efficient GPU optimization. A parallel GPU RRT provides route initialization when needed, and accelerated robot computations reduce the cost of refinement. Differentiable inverse dynamics incorporates payload-dependent torque limits and supports runtime inertial updates as carried objects change.
Time Optimization
TimeOpt optimizes traversal speed along the resulting fixed path.
Faster Iterations
Accelerated robot computations reduce the cost of refinement.
The Planning Problem
Planning Requirements
A planner must avoid obstacles, respect joint and torque limits, and still reach the goal in the shortest time.
RRT: Grow Branches to Find a Path
Sampling planners like RRT find a path in under a millisecond, but the path still needs time parameterization.
VAMP measured on an AMD Ryzen 9 7950X CPU (one thread).
TrajOpt (cuRoboV1): Smooth Waypoints
Trajectory optimization, as in cuRoboV1, refines waypoints toward optimal paths, but every waypoint keeps the same fixed time step.
B-Spline: C-Space Geometry
We instead optimize a B-spline: a few control points define a smooth path. After 25 iterations, 93 percent of queries succeed; cuRoboV1 needs five times as many.
Convergence comparison on an RTX 5090 GPU. The planning configuration uses 50 initial iterations followed by one 25-iteration refinement pass.
B-Spline: Time Optimization (Time Opt)
TimeOpt optimizes traversal time along the fixed path. A B-spline represents squared path speed, allowing time intervals to vary while enforcing sampled velocity, acceleration, jerk, and torque constraints.
Torque Limits With Heavy Payloads
At five kilograms, torque limits matter: cuRoboV2-Fast succeeds on 93 percent of queries, versus 78 for cuRoboV1.
Measured on an RTX 5090 GPU.
The Compute Problem
GPU RRT: Fast Seeds for Hard Problems
For hard problems, our GPU RRT finds seeds in 0.3 milliseconds, the fastest on a GPU.
Measured on an RTX 5090 GPU.
Inside One TrajOpt Iteration
Each iteration evaluates the spline, runs kinematics and inverse dynamics on parallel GPU streams, evaluates costs, and gathers gradients back to the spline controls for the L-BFGS update.
Where Iteration Time Goes
Kernel profiling identifies the main computation costs: kinematics, inverse dynamics, collision checking, and cost reductions.
Iteration Time After Kernel Changes
Faster computations, parallel execution, and fewer optimizer iterations reduce total optimization time. Reusing the cost at the current solution also avoids evaluating it again during line search.
Fast kernel profiling uses 100 bookshelf_small_panda:0 solves at 100 g, with eight cuboids and 50+25 iterations. Kernel durations can overlap; elapsed iteration time is measured separately. Baselines retain their native workloads.
Full Planning Pipeline
We build a full motion planning pipeline on GPU: IK with a linear seed, RRT only when needed, B-spline trajectory optimization, Time Opt, and validation.
Planning Time Across Devices
At 3 kg, planning takes 4.43 ms on RTX 5090 and 25.99 ms on Jetson AGX Orin, 14.7× and 23.2× faster than cuRoboV2 on each device, respectively.
Mean full-planning times at 3 kg on RTX 5090 and Jetson AGX Orin.
The Compute Problem: Kernels
Efficient B-Spline Gradient Gathering
Spline evaluation is cheap: each control gathers its gradients with one warp reduction, without atomics.
① Kinematics Kernels
Kinematics computes local poses in parallel. The backward pass gathers a wrench per link and maps it to joint gradients, avoiding repeated walks along the chain for each sphere.
Kernel profiling on an RTX 5090 GPU.
② Inverse Dynamics Kernels
Inverse dynamics runs chains in parallel and maps torque-loss gradients directly to joint-state gradients. Payload mass, center of mass, and inertia can change without recompiling.
Kernel profiling on an RTX 5090 GPU.
③ Collision Kernels
Link bounds skip separated links, and swept bounds skip distant obstacles. Adjacent half-sweeps share a distance evaluation and combine their gradients before rotating them to world coordinates.
Kernel profiling on an RTX 5090 GPU.
④ Fused Cost and Reduction Kernels
Fused cost reduction and gradient accumulation remove separate pointwise and gradient-add launches. The spline backward pass gathers cost and dynamics gradients directly into control gradients.
Kernel profiling on an RTX 5090 GPU.
Pick & Place Planning
Pick-and-Place: Planning and Cycle Time
Each cube needs four motions in two planning calls. Planning is nearly 13 times faster than cuRoboV1, and cycle time almost a third shorter.
Simulated on an RTX 5090 GPU.
Pick-and-Place: Side by Side in Simulation
Side by side in simulation, cuRoboV2-Fast finishes two cubes in 7.7 seconds, versus 11.0 for cuRoboV1 and 13.8 for cuRoboV2.
Simulated on an RTX 5090 GPU.
Real Robot
System Setup
SceneFields provides depth, detections, and robot masks. GraspGenX proposes grasps. cuRoboV2-Fast combines IK, optional GPU RRT initialization, B-spline trajectory optimization, and TimeOpt to plan each grasp motion in 23 milliseconds.
Measured on an RTX 6000 Ada GPU during real-robot runs.
Real Robot: Faster Motions and Shorter Cycles
On the real robot, arm motions are 45 percent shorter and cycles 35 percent shorter than cuRoboV2. Planning is one percent of the cycle.
Measured on an RTX 6000 Ada GPU during real-robot runs.
Full Narrated Video
Deploying cuRoboV2-Fast
Table Clearing
The robot transfers seven objects from the table in one continuous run of one minute and twenty-two seconds.
Bin Picking
In bin picking, every grasp is planned around the bin walls and neighboring objects.
Regrasping After Failure
When a grasp fails, the robot replans and approaches again.
Reacting to Scene Changes
As obstacles move, the map updates and cuRoboV2-Fast replans around them.
Fast Reaction
A newly placed box is mapped, grasped, and transferred in under seven seconds.
Grasp Another Object on Failure
Transferring a Thin Tripod
Pick and Transfer
Retry Quickly on Failed Grasps
React and Grasp a Bottle
Re-Grasp a Wooden Box
Cluttered Pick and Transfer
Clutter: Pick 1
Clutter: Pick 2
Clutter: Pick 3
Dense Clutter: Pick 1
Dense Clutter: Pick 2
Dense Clutter: Pick 3
Dense Clutter: Pick 4
Dense Clutter: Pick 5
Dense Clutter: Pick 6
Pick from a Sparse Table
Pick an Object from a Bowl
The gallery runs above use an RTX 6000 Ada GPU.
Small Object Manipulation
Thin and Flexible Objects
BibTeX
@misc{sundaralingam2026curobov2fast,
title = {cuRoboV2-Fast: Fast Torque-Limited Collision-Free Motion Planning},
author = {Sundaralingam, Balakumar and Murali, Adithyavairavan and Birchfield, Stan},
year = {2026},
eprint = {2611.06913},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2611.06913}
}