PushAround: Collaborative Path Clearing via Physics-Informed Hybrid Search

Zili Tang; Zongyuan Li; Zishao Qiao; Meng Guo

cs.ROOct 26, 2026

PushAround: Collaborative Path Clearing via Physics-Informed Hybrid Search

Authors: Zili Tang, Zongyuan Li, Zishao Qiao, Meng Guo

Organizations: Peking University, Beijing 100871, China: School of Advanced Manufacturing and Robotics · Peking University, Beijing 100871, China: College of Engineering

Abstract

Large vehicles in cluttered environments are often blocked by movable obstacles, motivating mobile robot teams to clear traversable corridors. Navigation Among Movable Obstacles (NAMO) approaches often abstract obstacle displacements while neglecting robot dimensions, obstacle mass, contact geometry and forces, yielding physically infeasible plans. PushAround is a physics-informed hybrid search that couples geometric connectivity reasoning with short-horizon simulation to construct executable clearing plans. Repeated W-clearance connectivity graph (WCCG) queries expose frontier gaps requiring widening and rank them by estimated downstream costs. The search generates direct or recursive pushing candidates, selecting obstacles, directions, contacts and forces, and admits physics-evaluated candidates. Numerical experiments validate WCCG against standard connectivity checks and demonstrate scaling to denser clutter and teams of up to four robots, with robustness to simulated execution-side physical-parameter mismatch.

Figures & tables

Fig. 1: Collaborative path clearing in simulation ( Top ) and hardware ( Bottom ). Two robots push movable obstacles to create a $W$ –clear path for a large vehicle from start to goal.

Fig. 2: Overview of the proposed framework. Top: Scene inputs, WCCG construction, gap ranking, and sampling of pushing modes for breaking frontier gaps; Bottom-left: Sim-in-the-loop hybrid search with parallel validation in simulation; Bottom-right: Online execution and adaptation during path clearing.

Fig. 3: Overview of the proposed physics-informed hybrid search (PIHS): WCCG construction, gap ranking, mode generation and physics validation.

Fig. 4: Direct pushing (Top) moves a gap-side blocker along a gap-opening direction, whereas recursive pushing (Bottom) first clears movable blockers in the swept region and then returns to the root gap-opening task.

Fig. 5: Mode generation for a gap-opening task: (a) local reachability, (b) the reachable contact candidates after twist filtering, (c) a compatible initial contact tuple, (d) greedy replacement of individual contacts, (e) the force balance of the optimized push, and (f) rollout validation of the mode.

Fig. 6: Sampled push tasks are evaluated via parallel simulation, producing updated states for search expansion.

Fig. 7: Execution and planning alignment in Scenario 1 ( Left ) and Scenario 2 ( Right ). Top: simulated snapshots during path clearing. Bottom: corresponding WCCG overlays with the start face in blue. The executed pushes progressively expand the reachable $W$ –clearance region until the goal is connected.

	Validation		Runtime (ms)
Type	CC $=$ FR	FP/FN †	CC	FR	WCCG
Convex	100/100	0/0	1.23	282.9	0.86
Non-convex	98/100	0/0	1.51	204.1	1.30
Mixed	98/100	0/0	1.33	242.4	1.05
Total	296/300	0/0	1.36	243.1	1.07

TABLE I: WCCG connectivity validation and mean runtime.

Method / Variant	Succ.(%)	PT (s)	ET (s)	#Sims	#Pushes
DFS-WCCG	25.0	$>$ 100.0	51.2	$>$ 500	8.0
SL-Push (offline)	62.5	0.06	44.1	0.0	7.3
SL-Push (sim)	75.0	18.2	64.6	20.0	10.2
Rec-NAMO	37.5	13.3	42.4	0.0	7.8
PushAround (ours)	97.5	10.3	28.6	121.5	6.0
w/o gap-lookahead	82.5	16.9	35.2	178.3	7.2

TABLE II: Performance comparison and ablations on Scenario 1 over

40

randomized trials. a

Fig. 8: Failure cases by baselines Rec-NAMO, SL-Push and DFS-WCCG.

Case	Trials	Succ.(%)	$PT$ (s)			$ET$ (s)
Obstacle density (PushAround, $N=2$ , nominal)
$M=15$	10	100	2.5	$\pm$	0.9	37.1	$\pm$	21.0
$M=30$	10	100	11.7	$\pm$	3.7	100.9	$\pm$	34.7
$M=45$	10	100	53.4	$\pm$	24.5	272.7	$\pm$	97.4
Robot-team size ( $M=30$ , nominal; 10 trials per cell)
PushAround, $N=3$	10	100	30.0	$\pm$	12.7	97.3	$\pm$	42.1

TABLE III: Scalability and robustness.

Fig. 9: Scalability of the clearing process. Left: density scaling ( $M{=}45$ , two robots): snapshots during clearing ( top ) and WCCG overlays with the start face in blue ( bottom ). Right: team scaling ( $M{=}30$ , four robots): initial and final states ( top ) and one push before contact and during execution ( bottom ).

Fig. 10: Obstacle-density scaling (

8\times 8

10

seeds per density). Times are budget-capped (planning

800

s, execution

20{,}000

steps

=500

s; dotted).

Fig. 11: Real-world pushing experiments with execution–planning alignment (over $20$ runs, $2$ scenarios). Top: two robots pushing obstacles in a $5{\times}6\,\mathrm{m}$ workspace. Bottom: WCCG overlays with the start face in blue.

Motion Planning

Motion planning finds safe, feasible paths and trajectories for robots and vehicles amid geometric, dynamic, and task constraints.

See all 404 papers

SPAR: Search-trained Push Action Ranker for Efficient Navigation Among Movable Obstacles
Oct 20, 2026Dhruv Metha Ramesh, Tri-An Nguyen, Junshan Huang +3Robotic Object Pushing Robot Motion Planning
Navigate or Relocate? Planning Among Movable Obstacles in Unknown Environments
Sep 17, 2026Yuqing Zhang, Haoyu Zhu, Yiannis KantarosRobot Task Planning Motion Planning
Department of Electrical and Systems Engineering, Washington University in St. Louis, St. Louis, MO, USA
Manipulation Feasible Navigation Among Movable Obstacles with Discrete Contact Pushing
Sep 20, 2026Shaohu Wang, Aiguo Song, Yulong Yuan +3Robot Motion Planning Motion Planning
State Key Laboratory of Bioelectronics, Jiangsu Key Laboratory of Robot Perception and Control Technology, School of Instrument Science and Engineering, Southeast University, Nanjing 210096, China

Oct 20, 2026J/K move · Enter open · S save

SPAR: Search-trained Push Action Ranker for Efficient Navigation Among Movable Obstacles

Dhruv Metha Ramesh, Tri-An Nguyen, Junshan Huang, Jingjin Yu+2

Robotic Object Pushing Robot Motion Planning Robotic Manipulation Motion Planning

Open paper AskPDF arXiv

Navigation Among Movable Obstacles (NAMO) requires a robot to rearrange objects that block access to its goal. For a pushing robot, progress may depend on a setup push that enables a later opening without itself clearing a passage. The number of candidate sequences can grow exponentially with sequence length, making physics-based search costly. We introduce the Search-trained Push Action Ranker (SPAR), which learns to prioritize push primitives from simulated search outcomes. Searches over sequences of up to two pushes per object provide supervision that distinguishes immediate openings from setup pushes. A transformer trained with regression and ranking losses scores primitives using contact tokens conditioned on scene geometry and reachability regions. The same model guides simulator evaluations in SPAR Search and selects actions for direct execution in SPAR Policy. Both modes solve successive keyhole problems that open access to neighboring free-space regions within a NAMO planner. Across successful runs on

300

simulated two-push keyhole problems, SPAR Search reduces median simulator calls by

7.0\times

, leading to a

3.8\times

reduction in planning time relative to Random Search. Without retraining, the same ranker guides search on

600

keyhole problems which have no solution within two pushes. At a depth limit of five pushes, SPAR Search achieves

50.7\%

success within

100

simulator calls, compared with

20.2\%

for Random Search. These gains carry over to NAMO planning: on hard simulated scenes, SPAR Search achieves

47\%

success within

100

simulator calls, compared with

21\%

for Random Search. In real-robot experiments, SPAR Policy succeeds in

37/40

trials without a physics-based digital twin, compared with

33/40

for SPAR Search and

27/40

for Random Search.

Abstract

Motion Planning

Explore similar work

SPAR: Search-trained Push Action Ranker for Efficient Navigation Among Movable Obstacles

Navigate or Relocate? Planning Among Movable Obstacles in Unknown Environments

Manipulation Feasible Navigation Among Movable Obstacles with Discrete Contact Pushing