From Manipulability to Motion:
Learning Point-Cloud-Constrained Proposals for Narrow-Space Planning

PCMS-Connect

Submitted to ICRA 2027

Being closer to a goal does not mean a robot has room to move toward it. PCMS-Connect learns which local motions remain available and uses this information to guide motion planning in narrow spaces.

Diana 7 in a narrow rack, with schematic point-cloud and constrained manipulability surface overlays illustrating observe, predict, and extend. — Observe the obstacle point cloud, predict a constrained manipulability surface, and propose a local motion. Surface and point-cloud overlays are schematic.

Core idea

1. Describe available motion

Directional bounds capture how joint limits, the robot itself, and nearby obstacles constrain local motion.

2. Propose an extension

A compact learned predictor helps choose a direction and step length, balancing task progress with available motion.

3. Check and connect

Every proposed edge is geometrically checked before it extends the bidirectional RRT-Connect trees.

Simulation and hardware

Narrow-space planning in simulation

A thin-shelf scene from the Panda benchmark. This clip is one qualitative example; aggregate results are reported below. Source replay at 2.5× speed.

Real-robot deployment on Diana 7

Two excerpts showing motion through narrow rack openings. The proposal mechanism transfers to Diana 7 using a predictor trained for that robot. Source replay at 3× speed.

Results at a glance

95.0% planning success (171/180 trials), compared with 86.1% (155/180) for goal-distance guidance across four families of Panda benchmark scenes.

39.8% lower online planning time on the 79 paired common successes in a 90-trial serial subset.

The Diana 7 demonstrations are qualitative hardware results, separate from the Panda benchmark statistics. Learned bounds guide candidate generation; finite edges still require geometric validation. The paper will be made available here at a later date.