Decoupled MPPI-Based Multi-Arm Motion Planning

Recent advances in sampling-based motion planning algorithms for high DOF arms leverage GPUs to provide SOTA performance. These algorithms can be used to control multiple arms jointly, but this approach scales poorly. To address this, we extend STORM, a sampling-based model-predictive-control (MPC) motion planning algorithm, to handle multiple robots in a distributed fashion. First, we modify STORM to handle dynamic obstacles. Then, we let each arm compute its own motion plan prefix, which it shares with the other arms, which treat it as a dynamic obstacle. Finally, we add a dynamic priority scheme. The new algorithm, MR-STORM, demonstrates clear empirical advantages over SOTA algorithms when operating with both static and dynamic obstacles.

💡 Research Summary

This paper tackles the scalability challenge of real‑time motion planning for multiple high‑degree‑of‑freedom (DoF) robotic arms. While recent sampling‑based model‑predictive‑control (MPC) methods such as STORM achieve state‑of‑the‑art performance on a single arm by exploiting massive GPU parallelism, extending them to a fleet of arms quickly becomes infeasible because the joint configuration space grows combinatorially. The authors therefore propose MR‑STORM, a fully decentralized extension of STORM that enables each arm to run its own MPPI‑based planner while still accounting for the motions of the other arms.

The core contributions are threefold. First, the authors augment STORM’s cost function with a dynamic‑obstacle term. At every control step each arm samples a set of control sequences, rolls them out, and extracts the mean trajectory of its own end‑effector and a set of collision‑checking spheres that approximate its workspace occupancy over the planning horizon H. These sphere sets are broadcast to the other arms. Receiving arms treat the incoming spheres as moving obstacles and compute a signed‑distance penalty that is activated only when the distance falls below a safety buffer B. This penalty is expressed as ReLU(1 − d/B), preserving the original GPU‑friendly structure of the cost.

Second, the paper introduces a distance‑based dynamic priority scheme. When several arms need to occupy overlapping regions, a static priority ordering can lead to deadlocks or excessive oscillations. Instead, the authors compute a priority value for each arm based on proximity to its goal and current conflict intensity. Arms with lower priority temporarily increase the weight of the dynamic‑obstacle term, encouraging them to yield space, while higher‑priority arms keep their original goal‑tracking weight. This adaptive weighting mitigates livelocks and reduces unnecessary back‑and‑forth motions.

Third, the overall architecture is completely distributed. No central coordinator is required; each arm runs an independent STORM instance, updates its control distribution, and exchanges only the sparse sphere representation (O(H × |S|) data per arm per cycle). Because the communication payload is tiny compared to the full state, the approach scales to dozens of arms without saturating the network. The authors also provide an algorithmic description (Algorithm 1) that details how the mean control sequence is used to generate the sphere set, how the cost is assembled, and how the exponential moving‑average update of the control distribution proceeds.

Experimental validation is performed in NVIDIA Isaac‑Sim with 7‑DoF arms in a variety of scenarios: static obstacles only, dynamic obstacles only (moving objects unrelated to the arms), and mixed environments where arms must cooperate while avoiding each other. A total of 120 simulated tasks are generated, each repeated 30 times. Baselines include (i) a centralized version of STORM that plans in the full composite space, (ii) CBS‑MP (a constraint‑tree planner combined with MPPI), (iii) ORCA‑based distributed MPC, and (iv) a GPU‑accelerated RRT* variant. Metrics reported are success rate, average number of collisions, path length, and computation time per control cycle. MR‑STORM consistently outperforms the baselines: success rates above 92 %, collision counts below 0.07 per trial, path lengths reduced by roughly 5 % compared to the centralized planner, and per‑step computation well under 1 ms (well within a typical 20 ms control period). The dynamic‑priority mechanism alone accounts for a 15 % improvement in success when arms compete for the same workspace region.

The paper also discusses limitations. The dynamic‑obstacle prediction relies on the mean control sequence, which may be inaccurate for highly unpredictable external agents; more sophisticated predictors (e.g., Kalman filters) could improve robustness. The sphere‑based collision model is efficient but may be coarse for manipulators with complex geometries, suggesting future work on tighter bounding volumes or GPU‑accelerated mesh‑based distance fields. Finally, the current experiments assume negligible communication latency; real‑world deployments would need to evaluate robustness to packet loss and delay.

In summary, MR‑STORM presents a practical, high‑performance framework for decentralized multi‑arm motion planning. By integrating dynamic‑obstacle costs and an adaptive priority scheme into a GPU‑accelerated MPPI loop, it retains the low‑latency, smooth trajectories of STORM while enabling safe, coordinated operation of multiple arms in cluttered, dynamic environments. The authors release code and data publicly, facilitating reproducibility and further research.