A robotics preprint reports much broader feasible exploration for QP-projected samples than for unfiltered or low-pass-filtered samples in a single-joint diagnostic. PRISM uses quadratic programming, or QP, a constrained optimization step, to project joint-velocity trajectories onto position, velocity, acceleration, jerk and initial-velocity constraints before rollout. The diagnostic’s effective covariance, a measure of the spread of feasible samples, was 0.4716 for QP, versus 0.0045 for unfiltered sampling and 0.0106 for low-pass filtering. The paper describes the QP exploration radius as about 100 times the unfiltered baseline and 45 times the low-pass result.
The result came from a narrow test. It compared unfiltered Gaussian sampling, Butterworth low-pass filtering and QP projection using 200,000 samples for each strategy. The violation threshold was below 0.3%, and the binary-search tolerance was 10^-4. The figures describe sampling behavior in that test, not direct manipulation success.
Task results varied by object and batch
The task-level evaluation covered four bimanual tasks: moving a tray, lifting and transporting a ball, handing over a cube and lifting a box. Goal poses were randomized, with 20 runs per task. The researchers recorded success rate, completion time for successful episodes and computation time for each planning step.
In a batch-size validation, ball-lifting success was 65% with a batch of 500 samples and nearly 100% with batches of at least 1,250. Tray-moving success stayed around 85% to 95%, handover reached up to 95%, and box lifting remained around 55% to 60%.
At the largest batch sizes, computation stayed below about 80 milliseconds for ball, tray and handover tasks and reached up to 130 milliseconds for box lifting. Completion times were reported only for successful episodes: handover took about 85 to 95 seconds and box lifting about 8 to 15 seconds.
Offline tuning was part of the evaluation
PRISM also included offline Bayesian optimization, an automated search over task-cost weights. The reported budget was 1,000 episodes per task, with a full run taking three to nine hours on an RTX 4090. Relative to manual tuning, the reported BO-weight results were 75% to 100% for ball lifting, 60% to 95% for handover, 45% to 90% for tray moving and 45% to 80% for box lifting. Average computation time remained comparable.
Human design remains part of the system. The authors state that the cost-function structure is manually designed, while its weights can be tuned offline. They also note that reliance on the MuJoCo model may expose sim-to-real gaps for deformable objects, fluids or highly dynamic contacts.
The comparison depended on the tested setup
The direct comparison covered baseline-CEM, low-pass filtering, Savitzky-Golay filtering and QP projection. On ball lifting, QP and low-pass filtering each recorded 95% success. QP recorded 90% for handover, versus 80% for low-pass filtering, 75% for Savitzky-Golay filtering and 15% for CEM. For tray moving, QP recorded 95%, compared with 70% for Savitzky-Golay filtering, 50% for low-pass filtering and 30% for CEM.
Box lifting reversed that pattern: low-pass filtering recorded 90% success, compared with 80% for QP, 70% for Savitzky-Golay filtering and 40% for CEM. The authors report comparable computation times across methods and negligible overhead from QP projection relative to the physics rollouts. In box lifting, the higher-success low-pass method also had longer successful-episode completion time.
The comparison has design limits. The baselines did not receive strict derivative bounds and required extra modifications for reliable task execution. The reported percentages came without confidence intervals or formal hypothesis tests, so they describe performance in the tested setup rather than a definitive ranking for other robots or workloads.
Physical evidence was reported, but narrowly
The paper reports operation at 10 to 16 hertz on a commodity laptop GPU and successful transfer from simulation to physical UR5e dual-arm manipulators. It does not report the number of physical trials or aggregate real-world success results, so this evidence is narrower than the simulation record.
In a separate preliminary obstacle-shift comparison using a Franka-Panda setup, PRISM succeeded in 60% of 20 trials, compared with 20% for the cited learned approach. The authors say the comparison was not definitive and used different assumptions.
A result bounded by its test conditions
The manuscript is an arXiv version 1 preprint dated 26 August 2026. It says code and supplementary videos are available through the project website. No funding source is named in the supplied text, and the acknowledgment says the material was produced exclusively by the authors and that AI was not used to generate figures or data.
Taken together, the study reports a broader-sampling result in a single-joint diagnostic, task results that varied across batch sizes and methods, and successful physical transfer in the demonstrated UR5e setting. The evidence remains bounded by the tested conditions: the covariance test was not a manipulation benchmark, physical trial totals are not reported, the cost structure remains manually designed and the learning-based comparison was preliminary.
Paper data and sources
Original title: PRISM: Projection-Integrated Sampling-Based MPC with Bayesian Cost Tuning for Bimanual Manipulation
Authors: Alinjar Dan, Iryna Hurova, Karl Kruusamäe, Arun Kumar Singh
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text