cs.RONov 4, 2026
SavecuRoboV2-Fast: Fast Torque-Limited Collision-Free Motion Planning
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 .
Figures & tables
Fig. 2: Our RRT evaluates multiple tree extensions in parallel, validates their edges separately, and connects the start and goal trees to recover a route. A parallel scan inserts accepted endpoints, retaining valid prefixes of blocked edges. Purple segments show unchecked proposals, while green segments have passed validation.
Fig. 3: Spline evaluation and backward gather: each sample uses four controls; gathers weighted sample gradients from the four orange segments and sums them using warp-level reduction primitives.
Fig. 4: TimeOpt schematic. (a) The position spline and its control points remain fixed; purple markers indicate positions at path knots. (b) Optimizing coefficients changes the squared path speed . (c) The resulting time map assigns unequal traversal times to the same path intervals. In this example, slower traversal through the shaded region and faster traversal elsewhere reduce the total duration.
Fig. 5: TimeOpt solver. The predictor and corrector reuse one Cholesky factorization. In the schematic, the corrector changes the purple direction to yellow; backtracking shortens the step to the green feasible update.
Fig. 6: Efficient kinematics: parallel local poses feed serial world composition, while prepared contributors enable independent Jacobian columns (top). The backward pass accumulates link wrenches once, then gathers them per actuator, avoiding repeated ancestor walks (bottom).
Fig. 7: Collision pruning and reuse. On-the-fly link AABBs eliminate sphere–sphere tests between separated links (left). Sweep bounds prune distant obstacles; adjacent half-sweeps share the SDF at and rotate their summed gradient once (right). denote neighboring midpoints.
Fig. 8: Efficient inverse dynamics. Forward RNEA uses compact spatial operations (top); backward computes loss gradients directly instead of forming full derivative matrices (bottom). Independent chains run concurrently, reducing at shared parents. Trees show leaf-to-root accumulation. Payload updates require no recompilation.
Fig. 9: Grasp-planning integration. IK selects a grasp, followed by pregrasp, approach, and lift motions in one planning call. Approach and lift follow Cartesian linear motions. A separate call plans transfer to the placement pose.
Fig. 10: Coarsely checked route initialization on MotionBenchMaker and MPINets (2,575 queries). Our RRT implementation is faster than pRRTC. VAMP plans routes faster, but VAMP + TOPPRA achieves lower motion-planning success with payloads than cuRoboV2-Fast (Fig. 12 ).
Fig. 11: Trajectory optimization convergence on 100 MotionBenchMaker cage_panda queries. Top: initial-only success. Bottom: cuRoboV2-Fast mean motion-time reduction on 93 shared valid queries after 25 initial iterations. Our B-spline formulation achieves higher success at small initial budgets. TimeOpt reduces mean motion time by 16% compared with finetuning using fixed timestep rescaling.
Fig. 12: Motion Planning with different payloads. cuRoboV2-Fast achieves similar success to cuRoboV2 while planning faster and producing shorter motions. Both V2 methods retain higher success than the TOPPRA baselines at the heaviest payload. Quality panels use RTX 5090 results; Fast success averages two repeats of 2,590 queries per payload. Motion duration and path length use 1,599 queries valid for all methods and payloads, including both Fast repeats. Separate planning-time panels show the 3 kg payload on RTX 5090 and Jetson AGX Orin, using different timing scales.
Fig. 13: Robot computation performance on RTX 5090. cuRoboV2-Fast accelerates robot computations and collision-aware IK. Latencies are medians; IK uses 100 shared Franka Panda targets. Other modules compare batches 1 and 1,024 on Franka Panda and G1. Broken-bar labels show full values; k denotes .
Fig. 14: Trajectory-optimization and full motion-planning times. Faster iterations and a smaller iteration budget compound to reduce optimization time. cuRoboV2-Fast achieves 11.0 and 10.5 shorter optimization-loop times than cuRoboV2 on RTX 5090 and Jetson AGX Orin, respectively. Iteration totals use 75 iterations for cuRoboV2-Fast and 300 for both baselines. Iteration timings on both devices use 100 repeated bookshelf_small_panda solves at 100 g.
Fig. 15: Mean per-cube planning time and planned motion duration for the five-cube simulated YAM tray arrangement. cuRoboV2-Fast plans over faster and produces shorter motions than both baselines.
Fig. 16: Simulated YAM transfers with 100 g (top) and 800 g (bottom) payloads. The heavier payload shifts motion from joint 2 to joint 5 to respect joint 2’s torque limit.
Fig. 17: Real-robot manipulation on one RTX 6000 Ada GPU. SceneFields provides depth, detections, and robot masks; its maps supply object surfaces to GraspGenX and collision distances to cuRoboV2-Fast. Our planner selects a grasp and plans collision-free, torque-limited pregrasp, approach, and lift motions in 23 ms. RRT initializes trajectory optimization when needed; transfer motions are planned separately.
Fig. 18: Arm utilization, motion time, and cycle time for real-robot object transfers. Utilization uses the most utilized joint at each instant. cuRoboV2-Fast nearly doubles time at 90–100% velocity utilization and reduces motion and cycle times by 44.8% and 34.6%, respectively.
Fig. 19: Real YAM table clearing and bin picking with cuRoboV2-Fast. Mean cycle breakdown over 53 completed sequences; planning accounts for only 1.0%, including failed planning calls and retries within each cycle. Other includes residual processing and waits.
More in this topic
Robot Motion Planning
Robot motion planning finds safe, feasible paths and trajectories for robots moving through constrained, changing environments.
See all 282 papers