In a tabletop shelf task, a seven-joint Franka Emika Panda arm moved a tomato to an upper shelf in 25 of 30 trials when its planner used SCOPE, researchers report. A comparison planner used the same model to propose paths but kept its search width fixed; it succeeded in 15 of 30 trials.
SCOPE, short for Score-Curvature for Online Precision Estimation, estimates how much each candidate path can vary at each point. The motion model keeps distinct routes separate; SCOPE estimates the room around each route instead of averaging different routes into one. The controller then searches more widely where a path is flexible and more narrowly in a tight passage. The boxes were repositioned between trials.
In the stationary-obstacle test, SCOPE’s controller collided in four of 30 trials and timed out without a collision once. Diff-MPPI, the comparison planner, collided in seven trials and timed out eight times.
In a separate test, a person reached toward a nearby Oreo box while the robot handled the tomato. With SCOPE predicting the hand’s motion, the robot succeeded in 21 of 30 trials, compared with 16 of 30 for MID-Ens-64, a controller that estimates uncertainty from 64 sampled trajectories. The SCOPE setup recorded three hand collisions, versus eight for MID-Ens-64. The paper notes that a trial could include both a hand collision and a collision with the surrounding scene.
On one NVIDIA A5000 GPU, the authors measured 12.87 milliseconds per step for the full SCOPE pipeline, including a final scaling adjustment learned from a separate set of trajectories. An ensemble of 64 sampled trajectories took 26.58 milliseconds in the same test; both figures are specific to this model and hardware.
The arXiv paper says it was accepted as a Spotlight presentation at CoRL 2026. The physical evaluation covers one robot and one shelf task, with 30 trials per method in each condition. The authors list extending the method to raw images or point clouds as future work. These trials do not establish safety or performance on other robots.
Image: A NIAID laboratory robot arm handles an assay plate. This illustrative image is not the Franka Panda arm or shelf used in the study.
Image courtesy of the National Institute of Allergy and Infectious Diseases.
Sources
Control-Ready Uncertainty for Trajectory Diffusion, arXiv; SCOPE project page