Abstract
Multi-robot-arm motion planning is a key challenge in deploying multiple manipulators for industrial tasks such as manufacturing. Existing search-based and sampling-based solvers often require significant computation time to produce collision-free, high-quality motions suitable for safe execution. In this work, we introduce a new suite of multi-robot-arm motion planners capable of near real-time motion generation, combining classical planning algorithms with state-of-the-art vectorized collision-checking techniques. Based on CPU SIMD instructions, our new planners remove motion validation as the primary bottleneck and achieve up to two orders of magnitude speedup in both motion planning and execution postprocessing for multi-arm manipulation tasks. We also release the implementation of our vector-accelerated multi-robot planning and execution algorithms, and we believe this will lower the barrier for research and development of multi-robot arm planning and manipulation problems.
Vectorized Multi-Robot Collision Checking
Figure 1: vectorized multi-robot collision checking achieves up to 148x motion-validation speedup.
VAMP-MR makes motion validation fast: CPU-SIMD vectorized forward kinematics and collision checking enable high-throughput evaluation over batches of multi-robot configurations.
Table 1 : Average collision checking time over 10,000 random samples. “Single” checks a single set of multi-robot configurations; “Motion” validates a motion segment between two configurations.
| Check | Environment | FCL (μs) | Ours (μs) | Speedup |
|---|---|---|---|---|
| Single | Panda Two Rod | 100.99 | 8.60 | 11.7× |
| Panda Four | 206.53 | 15.01 | 13.7× | |
| Panda Four Bins | 382.63 | 13.70 | 27.9× | |
| Motion | Panda Two Rod | 3936.13 | 36.03 | 109.2× |
| Panda Four | 4836.50 | 32.61 | 148.3× | |
| Panda Four Bins | 930.03 | 14.23 | 65.4× |
Evaluated environments: Panda Two Rod, Panda Four, Panda Four Bins.
Planning
Vectorized motion validation accelerates both composite RRT-Connect and CBS-style multi-robot planning across the benchmark environments by one to two orders of magnitude.
Shortcutting
Postprocessing is collision-checking heavy. Multi-robot-arm shortcutting with vectorized validation converges to high-quality solutions up to 50x faster than standard collision checking.
Example RRT planning + shortcutting rollout:
Panda Two Rod
Panda Four Bins
Panda Four
LEGO Assembly
We evaluate long-horizon dual-arm LEGO assembly tasks described in APEX-MR and measure runtime across task assignment, motion planning, and TPG construction.
Table 2: Final runtime and makespan for assembly planning, with and without vector acceleration averaged over 4 seeds.
| Metric | Cliff | Vessel | RSS | Chair |
|---|---|---|---|---|
| FCL-Based | ||||
| Total Time (s) | 12.1 | 32.2 | 71.5 | 845.2 |
| Final Makespan (s) | 185 | 397 | 741 | 2180 |
| Ours | ||||
| Total Time (s) | 3.39 | 8.38 | 31.2 | 59.3 |
| Final Makespan (s) | 159 | 340 | 542 | 2146 |
| # of Bricks | 11 | 36 | 47 | 258 |
Assembly executions (all videos are 10× speed):
Cliff (11 bricks)
Vessel (36 bricks)
RSS (47 bricks)
Chair (258 bricks)
Code
The main repository is vamp-mr/vamp-mr. Our codebase is built on top of vamp and APEX-MR. We support a set of prebuilt environments and an easy-to-use Python interface to define and run new multi-robot-arm environments from a portable JSON config.
Python Example
# Generate a new environment (compiles a C++ plugin for the new multi-robot-arm setup)
python3 mr_planner_core/scripts/plugins/generate_vamp_robot_plugin.py \
--env-name my_robot_env \
--base-transforms ${path_to_transforms}.json \
--robot-header /usr/local/include/vamp/robots/panda.hh \
--robot-struct Panda \
--num-robots 4 # Plan using the generated environment JSON
import mr_planner_core
env = mr_planner_core.VampEnvironment("my_robot_env.json")
res = env.plan(planner="composite_rrt", planning_time=5.0, shortcut_time=0.1, seed=1)
Citation
For now, please cite the project website based on the workshop version
@misc{vamp_mr_website,
title = {VAMP-MR: Vector-Accelerated Motion Planning and Execution for Multi-Robot-Arms},
author = {Huang, Philip and Gao, Chenrui and Li, Jiaoyang},
howpublished = {\url{https://github.com/vamp-mr/vamp-mr}},
note = {AAAI 2026 Workshop on Multi-Agent Path Finding (WoMAPF), project website},
year = {2026}
}
