Source-linked AI summary
CAMP: Cooperative Arm-Hand Motion Planning in Constrained Spaces
Ziyuan Wang, Yunlong Shan, Fei Mo, Sichao Liu, David Navarro-Alarcon, Jia Pan, Kosta Jovanovic, Xin Jiang, Peng Zhou
TL;DR
In constrained spaces, separate arm and hand planning can miss motions requiring coordinated adaptation, while full joint-space planning faces difficult high-dimensional search. CAMP combines arm-space guides, layered hand search with local arm relaxation, and VMP-based joint optimization; it reports high success across six simulated tasks and demonstrates planned motions on a physical robot.
Problem
Separate arm and hand planning can miss motions requiring coordinated adaptation, while direct search over the complete arm–hand configuration space becomes difficult as active joints increase.
Source: §I · p. 1Method
CAMP generates multiple arm-space guides, searches hand configurations in layers with local arm relaxation, and refines candidates through endpoint-preserving VMP-based joint optimization.
Source: §I · p. 2Results
Across six simulated tasks, CAMP achieves mean planning success rates of 84.2%–98.5%, and physical trials report task success rates of 80%, 90%, and 80% on three tasks.
Source: §V · p. 8, D. Real Robot Experiments · p. 8Takeaways & Limitations
Real-robot experiments validate the feasibility of planned motions on UR7e+LinkerHand for constrained manipulation tasks.
Source: §V · p. 8The evaluation is limited to one platform and static environments with known obstacle geometry, and execution is sensitive to obstacle localization, calibration, and robot execution errors in tight clearances.
Source: §V · p. 8
Abstract
from arXiv · showhide
Coordinated arm-hand motion planning is fundamental to dexterous robotic manipulation in complex and constrained environments. A straightforward solution is to decompose the problem into separate arm path planning and hand motion generation; however, this poses a dilemma: decomposition can miss feasible solutions that require coordinated arm-hand adaptation along the path. Alternatively, directly planning in the high-dimensional joint arm-hand configuration space captures such coupling but faces a substantially enlarged search space and nonconvex collision constraints. To characterize this coupling, we formulate feasible hand fibers that capture collision-free hand configurations for each arm configuration. Based on this formulation, we propose CAMP, a high-success and efficient cooperative arm-hand motion planner for constrained environments. CAMP constructs candidate trajectories through layered hand search with local arm relaxation, then compactly represents them using endpoint-preserving via-point movement primitives (VMPs) for coarse-to-fine joint optimization. Across six constrained simulation tasks, CAMP achieves 84.2-98.5% planning success, outperforming alternative planners with competitive efficiency. Ablation studies verify the contributions of arm relaxation, VMP representation, and coarse-to-fine optimization, while real-robot experiments demonstrate CAMP on constrained manipulation tasks. The project website is available at https://camp-armhand.github.io/.
I. INTRODUCTION
Dexterous manipulation in confined spaces depends on coordinating arm positioning with hand articulation, but planning these motions presents a trade-off between missed coupled solutions and difficult full-space search. CAMP uses configuration-dependent feasible hand fibers to characterize this coupling and jointly adapt arm and hand motion.
I. INTRODUCTION
- The Arm–Hand Planning Dilemma is that separate planning can miss motions requiring joint adaptation, while full arm–hand search becomes difficult as active joints increase.A collision-free arm path may also fail to admit a continuous collision-free hand motion between prescribed endpoints.Source: §I · p. 1
II. PROBLEM FORMULATION
- Configuration-dependent feasible hand fibers contain collision-free hand configurations for each arm configuration, and their shape and connectivity may change as the arm moves.A fixed hand shape can restrict the feasible arm set, and a fixed arm path can restrict feasible hand motion.Source: §I · p. 1, §II · p. 2
I. INTRODUCTION
- Prior work addresses grasp synthesis, arm-aware grasp generation, sampling-based planning, trajectory optimization, and simplified-space guidance, with distinct limits in trajectory planning or full-system feasibility.The cited review also notes posture-mapping requirements for one approach and arm-only planning in the reported setup of another.Source: §I · pp. 1–2
- CAMP generates multiple arm-space guides, uses layered hand search with local arm relaxation, and refines the resulting candidates through endpoint-preserving VMPs and coarse-to-fine joint optimization.Both arm and hand trajectories remain adjustable during optimization.Source: §I · p. 2
- Across six constrained simulation tasks, CAMP achieves 84.2–98.5% planning success, and the paper reports competitive efficiency and validation on real-robot tasks.The contributions also report ablations verifying key components.Source: §I · p. 2
III. METHODOLOGY
The methodology section outlines CAMP’s candidate-generation and optimization stages, from layered hand search to cooperative trajectory refinement.
III. METHODOLOGY
- Layered hand search with local arm relaxation is one component of the methodology presented for CAMP.The section overview identifies this as part of the candidate-generation process.Source: §III · p. 3
- CAMP represents candidate trajectories with VMPs as part of its methodology.The section overview identifies VMP trajectory representation as a method component.Source: §III · p. 3
- The methodology includes coarse-to-fine joint optimization of multiple candidates.The section overview presents this as a stage following the candidate-generation components.Source: §III · p. 3
A. Arm-Relaxed Layered Hand-RRT Initialization
Initialization combines arm-space reference paths with layered hand search, relaxing the arm locally when a fixed guide blocks a feasible hand connection.
A. Arm-Relaxed Layered Hand-RRT Initialization
- RRT-Connect generates candidate reference paths in arm configuration space between the prescribed start and goal configurations.The paths provide arm references for the subsequent hand search.Source: A. Arm-Relaxed Layered Hand-RRT… · p. 3
- Each arm reference path is divided into ordered progress layers, and bidirectional RRT searches for collision-free hand configurations connected across adjacent layers.The search initializes trees at the prescribed initial and goal hand configurations and applies a step-size bound to adjacent-layer connections.Source: A. Arm-Relaxed Layered Hand-RRT… · p. 3
- When an interpolated midpoint between layers is in collision, local replanning jointly adjusts hand configurations and arm offsets to explore nearby feasible fibers.The componentwise arm-offset interval bounds deviations from the reference configurations.Source: A. Arm-Relaxed Layered Hand-RRT… · p. 3
- Locally replanned segments are spliced into the original path to yield a complete arm–hand trajectory.Original path nodes outside the local window serve as connection anchors, and the search checks connections to the original path.Source: A. Arm-Relaxed Layered Hand-RRT… · p. 3
B. VMP Representation
CAMP encodes complete arm–hand candidates as endpoint-preserving VMP trajectories, reducing optimization to smooth basis weights. The representation supports different temporal resolutions for arm and hand motion and is initialized by regularized fitting.
B. VMP Representation
- VMPs represent complete arm–hand candidate trajectories using phase basis functions and weights.The arm and hand trajectories are decoded from the candidate using their respective VMP weights.Source: B. VMP Representation · p. 3
- The endpoint envelope vanishes at both ends, so VMP deformation preserves the trajectory’s start and goal configurations.The linear endpoint reference and normalized Gaussian basis functions define the trajectory representation.Source: B. VMP Representation · p. 3
- Smooth basis functions constrain deformation to a low-dimensional temporal subspace and couple neighboring waypoint changes, reducing dimension when NVMP ≪ Njoint.Choosing Ka > Kh allows finer temporal detail for local arm obstacle avoidance and a more compact basis for hand-shape changes.Source: B. VMP Representation · p. 4
- Regularized fitting of endpoint-preserving, resampled candidates initializes the arm and hand VMP weights.The fitting objective penalizes reconstruction error, weight magnitude, and second-order trajectory variation, and is solved as a quadratic problem.Source: B. VMP Representation · p. 4
C. Arm–Hand Joint Trajectory Optimization
CAMP jointly optimizes arm and hand VMP weights under trajectory-quality, collision, and joint-limit constraints. Because the arm determines the sequence of feasible hand fibers, changing the arm and hand together adapts both the available sets and the selected hand configurations.
C. Arm–Hand Joint Trajectory Optimization
- The shared coarse- and fine-stage objective combines trajectory quality, deviation from initialized weights, and a soft environment-collision penalty.Both arm and hand weights remain adjustable in both stages; trajectory quality penalizes excessive joint variation and high-frequency oscillations.Source: C. Arm–Hand Joint Trajectory… · p. 4
- Environment- and self-collision constraints require each decoded hand configuration to lie in the feasible fiber induced by its decoded arm configuration.The optimization also constrains joint positions and velocities and penalizes trajectory states near or inside environmental obstacles.Source: C. Arm–Hand Joint Trajectory… · pp. 4–5
- Jointly changing arm and hand weights changes both the phase-indexed sequence of feasible hand fibers and the hand configurations selected within them.Fixing arm weights instead leaves the hand path to be found through a prescribed fiber sequence.Source: C. Arm–Hand Joint Trajectory… · p. 5
D. Coarse-to-Fine Multi-Candidate Lift Optimization
CAMP retains multiple initialized arm–hand candidates and optimizes them in two stages. Coarse optimization prioritizes feasibility; selected candidates then receive longer refinement with greater emphasis on trajectory quality.
D. Coarse-to-Fine Multi-Candidate Lift Optimization
- CAMP preserves each candidate’s reference trajectories and initial weights throughout optimization, selection, and decoding.The candidates represent different arm routes and their associated hand motions.Source: D. Coarse-to-Fine Multi-Candidate Lift… · p. 5
- Parallel coarse optimization jointly adjusts candidate weights with a smaller iteration budget, prioritizing collision and other constraint violations.A stronger penalty on deviations from initial arm weights favors hand-dominant corrections while permitting necessary arm adjustments.Source: D. Coarse-to-Fine Multi-Candidate Lift… · p. 5
- After coarse optimization, candidates are ranked by constraint satisfaction and trajectory quality, and the top Nf candidates are selected for fine optimization.Numerically invalid candidates are discarded, and candidates satisfying joint-limit, velocity-limit, environment-collision, and self-collision constraints receive priority.Source: D. Coarse-to-Fine Multi-Candidate Lift… · p. 5
- Fine optimization gives greater relative weight to trajectory quality, reduces the penalty on initial arm weights, and uses a larger iteration budget before the best-quality trajectory is returned.The fine stage continues to jointly adjust arm and hand weights.Source: D. Coarse-to-Fine Multi-Candidate Lift… · p. 5
IV. EXPERIMENTS
The experiments evaluate cooperative planning against alternatives, test CAMP’s design components, and assess real-robot execution. The supplied passages describe six simulated tasks, the platform and trial setup, and the evaluation metrics, but do not report comparative outcomes.
IV. EXPERIMENTS
- The experiments ask whether cooperation improves feasibility over decoupled planning, how CAMP compares with alternatives, how its components affect performance, and whether its trajectories execute on a real robot.The evaluation includes six simulated tasks and real-robot experiments.Source: §IV · p. 6
- The simulated platform uses a 6-DoF UR7e arm and 16-DoF LinkerHand, with exact final trajectory collision checks in MuJoCo.Each task–method batch contains 100 trials; trajectories use 200 waypoints, and CAMP refines the top Nf = 3 candidates.Source: §IV · p. 6
- Success Rate measures the percentage of trials producing a complete 22-DoF trajectory that meets task constraints, reaches the prescribed goal, and passes exact collision checks.Planning Time is wall-clock solver time, reported as the arithmetic mean over all 100 trials.Source: §IV · p. 6
- Normalized Joint-Space Trajectory Length measures cumulative joint variation relative to direct start–goal displacement; one denotes straight-line interpolation.N.T.L. is averaged over successful trajectories for each method.Source: §IV · p. 6
A. Does Arm-Hand Cooperative Planning Improve Feasibil-
CAMP achieves the highest success rate across all four tasks, where decoupled plans cannot adapt hand shape and arm motion together. In Cabinet Cylinder Pregrasp, coordinating reopening with arm motion permits passage through a barrier and approach to the cylinder.
A. Does Arm-Hand Cooperative Planning Improve Feasibil-
- CAMP achieves the highest S.R. across all four tasks.The comparison uses two decoupled baselines: one fixes the initial hand configuration, while the other plans with a task-dependent preshape.Source: ity over Decoupled Planning? · p. 6
- A fixed initial hand configuration cannot adapt to spatial constraints that vary along the path.A task-dependent preshape suits some passages, but cannot handle motions requiring hand-shape transitions.Source: ity over Decoupled Planning? · p. 6
- In Cabinet Cylinder Pregrasp, CAMP keeps the hand compact past the front barrier, then reopens near the cylinder while adjusting the arm pose for collision avoidance.Source: ity over Decoupled Planning? · p. 6
B. How Does CAMP Compare with Alternative Motion Planners?
CAMP consistently achieves the highest success rate in comparisons with alternative planners. The reported discussion links its performance to candidate generation that adapts arm and hand together, followed by joint trajectory refinement.
B. How Does CAMP Compare with Alternative Motion Planners?
- CAMP consistently achieves the highest S.R. among the compared motion planners.Table II compares planners over task–method pairs and reports S.R. and P.T. as means and standard deviations across batches; the supplied passages do not include the table's cell values.Source: B. How Does CAMP Compare with… · p. 6, Table II · p. 7
- QRRT* failures stem mainly from search timeouts and collisions, while lifting paths through successive robot abstractions is difficult under strong arm–hand coupling.CAMP uses bidirectional layered hand search on multiple arm reference paths and locally relaxes blocked transitions before joint VMP refinement.Source: B. How Does CAMP Compare with… · p. 7
- A* initialization improves CHOMP in several constrained tasks, but optimization baselines still rely on a single full-joint initialization and thus offer less route diversity.Source: B. How Does CAMP Compare with… · p. 7
- CAMP maintains high S.R. with competitive planning efficiency, supported by compact VMPs that reduce optimization dimension and coarse-to-fine refinement.The passage attributes the balance to reduced-dimensional global guidance and compact trajectory optimization.Source: B. How Does CAMP Compare with… · p. 7
C. Ablation Study of CAMP
The ablation experiments use the simulated Ball-in-Box Pregrasp task, with 100 planning trials per setting and other planning and validation settings fixed.
C. Ablation Study of CAMP
- Each ablation setting uses 100 Ball-in-Box Pregrasp trials, with other planning and validation settings held fixed.Source: C. Ablation Study of CAMP · p. 7
1) Comparison of Fixed-Arm and Arm-Relaxed Initialization:
Across six simulated tasks, CAMP reports mean planning success rates of 84.2%–98.5%, and real-robot trials demonstrate planned-motion feasibility on one platform in static environments with known obstacle geometry. The supplied ablation and robot passages describe the tested mechanisms, outcomes, and execution boundaries.
1) Comparison of Fixed-Arm and Arm-Relaxed Initialization:
- With arm relaxation disabled, 12 of 100 trials produced no more than two valid initial trajectories, and only one of those trials succeeded; CAMP produced at least two valid initial trajectories in every trial.A fixed arm guide restricts available hand motions between layers, while local arm relaxation can help generate complete candidates.Source: 1) Comparison of Fixed-Arm and… · p. 7
2) Comparison of Trajectory Representations:
- Under the tested setup, VMP uses 500 weights versus 4,356 waypoint-wise variables, and increases S.R., reduces P.T., and shortens mean N.T.L.The compact representation retains independent trajectory parameters for all active joints and supports continuous arm motion and hand-shape changes.Source: 2) Comparison of Trajectory… · pp. 7–8
3) Comparison of Coarse-to-Fine and Fine-Only Optimization:
- The coarse stage more strongly penalizes deviations from initial arm weights, favoring hand-shape corrections; the fine stage then relaxes that penalty and emphasizes trajectory quality.Direct fine-only optimization makes larger simultaneous arm-route and hand-shape changes from the outset, making local constraint satisfaction more difficult.Source: 3) Comparison of Coarse-to-Fine and… · p. 8
D. Real Robot Experiments
- Across 10 physical trials per task, success rates were 80% for Ball-in-Box Pregrasp, 90% for Display Button Press, and 80% for Cabinet Cylinder Pregrasp.Pregrasp success required reaching the goal pose, not grasping or lifting; button-press success required fingertip arrival, not a physical press. Most failures involved hand–obstacle collisions in constrained regions.Source: D. Real Robot Experiments · p. 8
V. CONCLUSIONS
- Across six simulated tasks, CAMP achieves mean planning success rates of 84.2%–98.5%; the evaluation is limited to one platform and static environments with known obstacle geometry.Real-robot experiments demonstrate feasibility on UR7e+LinkerHand; future work will incorporate modeling and execution uncertainty into collision constraints.Source: §V · p. 8
START GOAL
- Figure 5 shows robot snapshots over time from left to right; rows (a)–(c) depict Ball-in-Box Pregrasp, Display Button Press, and Cabinet Cylinder Pregrasp.Secondary views show hand details as the robot approaches constrained regions.Source: Figure 5 · p. 9