Source-linked AI summary
SRMP: Search-Based Robot Motion Planning Library
Itamar Mishani, Yorai Shaoul, Ramkumar Natarajan, Jiaoyang Li, Maxim Likhachev
TL;DR
Sampling-based planners can produce highly variable solutions, while existing tools may yield unpredictable motions, inconsistent datasets, and variable execution results. SRMP is a search-based manipulation-planning framework that supports single- and multi-robot planning and integrates with MoveIt!, simulators, and hardware. SRMP was more consistent and almost always lower-cost than OMPL in single-robot problems; xECBS averaged 1.5 seconds for Binpick and 2.41 seconds for Shelves, versus OMPL's 0.1 and 0.14 seconds.
Problem
Sampling-based planners can produce highly variable solutions, while existing tools may yield unpredictable motions, inconsistent datasets, and variable execution results.
Source: §I · pp. 1–2Method
SRMP is a search-based manipulation-planning framework that supports single- and multi-robot planning and integrates with MoveIt!, simulators, and hardware.
Source: §I · p. 2Results
SRMP was more consistent and almost always lower-cost than OMPL in single-robot problems; xECBS averaged 1.5 seconds for Binpick and 2.41 seconds for Shelves, versus OMPL's 0.1 and 0.14 seconds.
Source: D. Results · pp. 4–5Takeaways & Limitations
SRMP offers a framework for predictable manipulation planning, including multi-robot tasks, with the authors reporting rigorous guarantees and low-cost motions.
Source: §V · p. 7
Abstract
from arXiv · showhide
Motion planning is a critical component in any robotic system. Over the years, powerful tools like the Open Motion Planning Library (OMPL) have been developed, offering numerous motion planning algorithms. However, existing frameworks often struggle to deliver the level of predictability and repeatability demanded by high-stakes applications -- ranging from ensuring safety in industrial environments to the creation of high-quality motion datasets for robot learning. Complementing existing tools, we introduce SRMP (Search-based Robot Motion Planning), a new software framework tailored for robotic manipulation. SRMP distinguishes itself by generating consistent and reliable trajectories, and is the first software tool to offer motion planning algorithms for multi-robot manipulation tasks. SRMP easily integrates with major simulators, including MuJoCo, Sapien, Genesis, and PyBullet via a Python and C++ API. SRMP includes a dedicated MoveIt! plugin that enables immediate deployment on robot hardware and seamless integration with existing pipelines. Through extensive evaluations, we demonstrate in this paper that SRMP not only meets the rigorous demands of industrial and safety-critical applications but also sets a new standard for consistency in motion planning across diverse robotic systems. Visit srmp.readthedocs.io for SRMP documentation and tutorials.
I. INTRODUCTION
The introduction identifies a need for reliable, repeatable motion planning as robotics becomes more accessible and existing sampling-based planners can produce variable motions. It presents SRMP as a search-based manipulation framework with broad robot, simulator, and software support.
I. INTRODUCTION
- Sampling-based planners such as OMPL can produce highly variable solutions, while existing tools may yield unpredictable motions, inconsistent datasets, and variable execution results.These properties are especially problematic where motion consistency and interpretability matter.Source: §I · pp. 1–2
- Reliable, repeatable plans matter in shared industrial workspaces, and motion planners are also used in robot-learning systems.Workers need to anticipate robot motions, while learning approaches can rely on planners for demonstrations or movement between key poses.Source: §I · pp. 1–2
- Fig. 2 compares consistency, solution costs, and planning times across repeated bin-picking and shelf-to-shelf planning problems.Each of 15 problems per scenario was repeated 10 times with slight goal-configuration variations; two anytime algorithms had a five-second limit.Source: Figure 2 · p. 2
- SRMP uses graph-search algorithms with completeness and bounded-suboptimality guarantees to produce short, repeatable motions for manipulation.The framework is designed for consistent, safe, and reliable motion generation.Source: §I · p. 2
- The introduction frames SRMP as a framework for modern robotics research and practice, including data collection and industrial automation.It links this scope to accessible simulators and demand for powerful, reliable, integrable planning libraries.Source: §I · pp. 1–2
- SRMP supports single- and multi-robot planning and integrates with MoveIt!, simulators, and real hardware through C++ and Python interfaces.The paper describes it as the first motion-planning library to support both single-robot and multi-robot manipulation planning.Source: Figure 1 · p. 1, §I · p. 2
II. SRMP
SRMP is a search-based motion-planning library for robotic manipulation. It is designed to work standalone or plug into existing pipelines, with a C++ implementation and Python bindings.
II. SRMP
- SRMP provides search-based algorithms for solving robotic-manipulation motion-planning problems.Source: §II · p. 2
- The standalone library can be integrated into existing pipelines, including MoveIt! as a plugin.Source: §II · p. 2
- SRMP is implemented in C++ for performance and provides Python bindings for ease of use.Source: §II · p. 2
A. Usage
SRMP can be used as a standalone library or as a MoveIt! plugin. Its interface configures a planning world and returns trajectories, including for coordinated multi-robot planning.
A. Usage
- The PlannerInterface loads robot models, manages environment objects, configures start and goal conditions, and returns trajectories with configurations, velocities, accelerations, and timestamps.Source: 1) Standalone Library · p. 2, A. Usage · p. 2
1) Standalone Library:
- In standalone use, Python code loads a robot and environment, specifies a start and goal, then computes a trajectory; multi-agent planning can plan for both robots.The library supports manipulators planned independently or as a coordinated system, with collision checking via COAL and kinematics via Pinocchio.Source: 1) Standalone Library · pp. 2–3, Figure 3 · p. 3
2) MoveIt Plugin:
- The MoveIt! planning adapter forwards requests to SRMP and returns its computed trajectory to MoveIt! for execution.MoveIt! supplies the kinematics and collision-checking modules used by the plugin.Source: 2) MoveIt Plugin · p. 3
III. EXPERIMENTAL ANALYSIS
The experimental analysis compares SRMP with OMPL on manipulation and multi-arm motion-planning problems. The authors report more consistent, repeatable solutions at comparable planning times, with support for multi-robot planning.
III. EXPERIMENTAL ANALYSIS
- The analysis covers a variety of manipulation and multi-arm motion-planning problems.Source: §III · p. 3
- SRMP is compared with OMPL, described as the de-facto standard motion-planning library.Source: §III · p. 3
- The authors report that SRMP produces more consistent and repeatable solutions in comparable planning times while scaling to multi-robot planning problems.Source: §III · p. 3
A. Experimental Setup
The evaluation repeats task-relevant planning problems with slight goal perturbations to test consistency, and assesses path cost, consistency, planning time, and success. It also measures effective runtime ratios for multi-robot tests.
A. Experimental Setup
- 10 single-robot and 5 multi-robot repetitions per problem use slight goal perturbations to test whether similar queries yield consistent solutions.Start and goal configurations are drawn from collision-free, task-relevant configurations.Source: A. Experimental Setup · p. 3
B. Evaluation Metrics
- Path cost is end-effector travel distance in meters, summed across robots for multi-robot problems.The paper uses this measure to assess trajectory efficiency and as an indicator of how interpretable solutions are.Source: B. Evaluation Metrics · p. 4
- Consistency is the average coefficient of variation (CV) of path costs across repetitions; lower CV means costs are closer to their mean.CV is standard deviation divided by mean; a high CV indicates paths of varying quality.Source: B. Evaluation Metrics · p. 4
- The evaluation also reports average planning time and success rate; multi-robot tests include effective runtime ratios against the fastest planner for each problem.Single-robot and multi-robot time limits are 5 and 20 seconds, respectively.Source: B. Evaluation Metrics · p. 4
C. Evaluated Algorithms
The evaluation compares SRMP search planners with OMPL sampling-based planners in single- and multi-robot manipulation. SRMP reports lower-cost, more consistent solutions overall, with multi-robot xECBS finding solutions for almost all problems.
C. Evaluated Algorithms
- Single-robot comparisons pair SRMP MHA*, PA*SE, wA*, and ARA* with OMPL planners; multi-robot tests compare xECBS with BiTRRT, PRM, and RRT-Connect.The single-robot evaluation also includes CHOMP; multi-robot tests cover up to four robots, though SRMP can scale further.Source: C. Evaluated Algorithms · p. 4
D. Results
- In single-robot problems, SRMP algorithms were significantly more consistent and almost always had lower solution costs at comparable planning times.ARA* recovered optimal paths in some cases by terminating early, while RRT* ran to its runtime limit.Source: D. Results · p. 4
- In multi-robot problems, xECBS produced shorter paths and solved almost all problems, while OMPL algorithms failed on at least 20%.Average planning times were 1.5 seconds for Binpick and 2.41 seconds for Shelves with xECBS, versus 0.1 and 0.14 seconds for OMPL, respectively.Source: D. Results · pp. 4–5
- Across domains and embodiments, SRMP plans had consistently lower solution costs than OMPL, slightly slower average planning times, and perfect success rates.The stress tests used the Motion Benchmarker; search heuristics included joint-space Euclidean distance and backward task-space BFS from the goal.Source: D. Results · p. 5
IV. BACKGROUND
Motion planning finds a cost-minimizing trajectory from a start configuration to a state meeting the goal condition while remaining collision-free. Multi-robot planning combines the robots’ configuration spaces and requires avoiding inter-robot collisions.
IV. BACKGROUND
- Articulated manipulators commonly have 6 or 7 degrees of freedom, while multi-robot configuration-space dimensionality grows exponentially with the number of robots.The paper notes that this growth makes many conventional algorithms impractical at scale.Source: §IV · p. 5
A. Motion Planning Problem
- Motion planning seeks a trajectory from a start state to a goal-satisfying state that minimizes a cost function.The problem is formulated in configuration space; for manipulation, this may be the space of joint angles.Source: A. Motion Planning Problem · p. 5
- For multi-robot manipulation, the configuration space is the Cartesian product of each robot’s configuration space, and the goal condition checks all robots.The robots must also avoid colliding with one another.Source: A. Motion Planning Problem · p. 5
B. Algorithms Taxonomy
Motion-planning methods are grouped as sampling-, optimization-, inference-, or search-based. Their trade-offs include sampling variability and scalability, optimization guarantees and initialization, demonstration dependence, and search determinism and completeness.
B. Algorithms Taxonomy
- Sampling-based planners use random continuous samples to build a graph or tree and find a path linking start and goal states.They explore high-dimensional spaces quickly, but randomness can yield very different solutions to nearly identical problems and may require many samples for short paths.Source: B. Algorithms Taxonomy · p. 6
- In multi-agent planning, sampling in the composite configuration space raises computational complexity and reduces scalability for short-path planning.The composite space is formed from the Cartesian product of individual configuration spaces.Source: B. Algorithms Taxonomy · p. 6
- Optimization-based methods seek locally optimal trajectories using numerical optimization, but generally lack completeness, optimality, or convergence guarantees for nonlinear systems.Collision avoidance makes manipulation planning non-convex, and GCS methods require laborious preprocessing that restricts use to fixed settings with infrequent scene changes.Source: B. Algorithms Taxonomy · p. 6
- Inference-based methods use probabilistic models to infer likely trajectories, often from demonstrations, but depend on demonstration quality and quantity and may not generalize to unseen scenarios.Collecting demonstrations at scale is especially challenging for multi-agent coordination.Source: B. Algorithms Taxonomy · p. 6
- Search-based methods systematically explore configuration space and are deterministic and complete, often with global-optimality or bounded-suboptimality guarantees.They build graphs implicitly, use heuristics and motion primitives, and have been applied to multi-robot planning at scale.Source: B. Algorithms Taxonomy · p. 6
C. Search-Based Planning Algorithms in SRMP
SRMP’s algorithms use prioritized state-space search over implicit graphs, with methods that improve prioritization, anytime refinement, and parallel speed. For multi-robot manipulation, xECBS decouples arm planning while preserving completeness and bounded sub-optimality guarantees.
C. Search-Based Planning Algorithms in SRMP
- SRMP algorithms explore states through a priority queue, with priority functions that can guarantee optimality when consistent or admissible.Generated but unexpanded states remain in the queue and are expanded according to their priority.Source: C. Search-Based Planning Algorithms in… · p. 7
- Implicit graphs built from domain motion primitives reduce memory use by storing only search-relevant portions of otherwise prohibitively large state spaces.Manipulation planning’s high dimensionality makes explicit graph representation impractical.Source: C. Search-Based Planning Algorithms in… · p. 7
- MHA* uses inadmissible heuristics with bounded sub-optimality guarantees, while ARA* finds an initial solution before refining it within a time limit.Parallel methods such as wPA*SE accelerate search across multiple cores while preserving bounded sub-optimality guarantees.Source: C. Search-Based Planning Algorithms in… · p. 7
- xECBS addresses multi-robot state-space growth by decoupling planning across arms while guaranteeing completeness and bounded sub-optimality.It finds collision-free solutions by planning for individual arms and imposing constraints.Source: C. Search-Based Planning Algorithms in… · p. 7
D. Existing Software Packages
Existing motion-planning packages offer varied capabilities, but limitations in predictability, scalability, integration, or maintenance leave gaps for practitioners. The cited packages do not provide native multi-arm planning support.
D. Existing Software Packages
- OMPL offers extensive sampling-based algorithms and broad integration, but reliability and predictability remain critical needs in industrial applications and robot learning.These needs relate to safety around people and reproducibility of training data.Source: D. Existing Software Packages · p. 7
- Designing effective admissible priority functions remains a significant challenge for search-based planning algorithms.The footnote qualifies the use of priority functions associated with optimality guarantees.Source: D. Existing Software Packages · p. 7
- ROS Motion Planning does not support high-dimensional planning, while optimization-based libraries may require additional integration effort.CuRobo offers computational advantages but requires CUDA-capable GPUs and can be challenging to integrate; SBPL is no longer actively maintained and lacks modern-framework support.Source: D. Existing Software Packages · p. 7
- None of the discussed libraries provide native multi-arm planning support, despite their strengths in other use cases.The passage identifies this as a critical gap in existing motion-planning frameworks.Source: D. Existing Software Packages · p. 7
V. CONCLUSION
The conclusion presents SRMP as software for reliable, consistent planning in single- and multi-robot manipulation. It emphasizes accessible interfaces, deployment options, and its claimed novelty in efficiently solving multi-robot planning with guarantees and low-cost motions.
V. CONCLUSION
- SRMP is designed to produce consistent plans for single- and multi-robot manipulation systems.The authors describe it as reliable, predictable, and consistent.Source: §V · p. 7
- SRMP supports Python and C++, and can run standalone or as a MoveIt! plugin.The conclusion also reports demonstrated effectiveness across a variety of scenarios and includes example code.Source: §V · p. 7
- The authors claim SRMP is the first framework, to their knowledge, to efficiently solve multi-robot manipulation problems with rigorous guarantees and low-cost motions.Source: §V · p. 7