PCMS-Connect · Submitted to ICRA 2027

From Manipulability
to Motion.

Learning Point-Cloud-Constrained Proposals
for Narrow-Space Planning

Watch the videoRun the codeExplore the results
Diana 7 reaching into a wheeled rack: obstacle points constrain the manipulability surface and a selected direction extends toward the red-block goal.

Directional motion bounds guide both the direction and length of local joint-space proposals.
Conceptual surface overlays illustrate the mechanism on the physical setup.

95.0%Planning success171/180 eligible Panda trials; goal-distance guidance: 155/180 (86.1%).
39.8%Lower online planning timeAgainst goal-distance guidance on 79 paired serial trials solved by both methods.
2 robotsSimulation to hardwarePanda benchmark evaluation and qualitative Diana 7 deployment with a robot-specific model.
01 / In motion

See the planner at work.

Five simulated planning cases followed by physical rack-reaching experiments on Diana 7.

Simulation comparisons use the same scene and goal for all three methods.Open video in a new tab ↗
02 / The mechanism

Local capability becomes a search edge.

The Point-Cloud-Constrained Manipulability Surface describes directional motion bounds at the current robot configuration.

01 — OBSERVE

State + obstacle points

Combine the current joint configuration with nearby obstacle observations.

02 — PREDICT

Directional bounds

A shared predictor approximates constrained radial bounds along joint-velocity rays.

03 — PROPOSE

Direction + step length

Combine available motion with task progress to construct finite joint-space candidates.

04 — EXTEND

Validate + insert

Geometrically validate candidate edges before extending the bidirectional search trees.

03 / Panda benchmark

Three planners. The same task.

Matched local RRT-Connect, goal-distance guidance, and PCMS-Connect, shown alongside the predicted surface at the current pose.

Animated three-way comparison for the thin-shelf Panda task, including the changing predicted surface.

Reaching into a thin shelf

The lower panel follows changes in the learned directional bounds. The arrow indicates the replayed motion direction. Shown clips are selected examples; aggregate results are reported above.

Another view: navigating a narrow passage
Animated comparison of three planners traversing a narrow passage.

Simulation footage replays saved joint paths. The displayed run time combines recorded online planning time and velocity-retimed path replay; it is not a measured hardware execution time.

04 / Diana 7 hardware

The same mechanism, a second robot.

Qualitative experiments use RealSense D435i point clouds and one fixed, robot-specific trained model across the tested rack configurations.

Diana 7 physically reaching toward a red block inside a wheeled rack.

Reach into the rack

The red block defines the goal. Experiments vary the rack placement and the goal position.

Two Diana 7 executions with the rack flipped; the right-hand setup is also lower.

Flip the rack. Reduce the clearance.

Flipping the rack creates a narrower opening. The right-hand configuration also lowers the rack height.

These demonstrations support qualitative cross-robot deployment of the proposal mechanism. Hardware demonstrations are separate from the 180-trial Panda benchmark statistics.
05 / Try it

Pretrained models. A small runnable release.

The code package includes model weights, four Panda scene families, the shared point-cloud collision interface, and a path viewer.

Python 3.10+ · CPU inference · No training required

Installation & running guide
python -m pip install -r requirements.txt

python run.py --method ours
python run.py --method goal_distance
python run.py --method rrt_connect