Organizations: 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
In environments with large movable obstacles, detour-only navigation can be inefficient or even infeasible, while obstacle interaction requires reasoning about navigation benefit, feasible placement, and executable manipulation. We present a hierarchical navigation among movable obstacles (NAMO) framework for mobile manipulators. At the high level, the planner identifies key blocking obstacles from reference paths and searches for relocation plans that jointly satisfy geometric, manipulation, and downstream navigation constraints. When direct relocation is hindered by other movable objects, a large language model (LLM) is selectively invoked to infer auxiliary manipulation dependencies, which are then verified by deterministic geometric planning. To execute the resulting relocation goals, we define discrete contact modes on the surfaces of box-shaped obstacles and select contact faces and regions online based on position and orientation errors, enabling straight, side, and corner pushing through contact switching. A recurrent reinforcement-learning policy coordinates the mobile base and manipulator to track tool center point (TCP) targets while preserving end-effector reachability during sustained pushing. Simulation and real-robot experiments demonstrate feasible navigation-manipulation in detour, single- and multi-obstacle relocation, and dependency-constrained scenarios, validating the framework for interactive navigation with large non-graspable obstacles. The open-source project is available at https://cloudytosunny.github.io/NAMO_DCPushing/.
Figures & tables
Fig. 1: Illustration of manipulation-feasible NAMO with discrete contact modes. Straight and side pushing skills are composed from the discrete contact set. The green curve denotes the navigation-only route, while the blue curve denotes the NAMO route, where the obstacle is pushed to a planned placement to restore traversability. The proposed framework selects the lower-cost route for efficient goal reaching.
Fig. 2: Overview of the proposed hierarchical NAMO framework. The planning layer evaluates navigation-only and NAMO hypotheses and selects the feasible plan with the lowest total cost. NAMO planning identifies key blockers and searches for manipulation-feasible relocation poses; when direct relocation fails, an LLM is selectively invoked to infer auxiliary obstacles and their manipulation dependencies for replanning. The execution layer maps relocation goals to discrete contact modes and TCP reference targets, which are tracked by a learned whole-body controller to compose straight, side, and corner pushing, while navigation-only segments follow the selected collision-free path.
Fig. 3: Robot–obstacle interaction with discrete contact modes. (a) Each vertical box face contains left, center, and right contact anchors, which define TCP reference targets and induce different translational and rotational effects. (b) In Stage 1, the pushing direction is aligned with the target before straight pushing. (c) In Stage 2, contact selection jointly considers position and orientation errors to reach the target pose.
Fig. 4: Planning results in four representative scenarios. Green denotes the selected route and obstacle relocation poses. (a) and (b) contain no feasible navigation-only path, whereas (c) and (d) show different route selections caused by small environmental changes.
Fig. 5: Planning and execution results in dependency-constrained scenarios. (a) and (b) show LLM-assisted planning, where o2 and o3 are key blockers and o1 and o4 are auxiliary obstacles, with o1≺o2 and o4≺o3 . (c) and (d) show the simulated execution, in which the planned relocations restore connectivity and enable the robot to reach the goal.
Simulation scene
Strategy
Jnav
Jman
Jtotal
1
NAMO
23.59
5.85
29.44
2
NAMO-A
15.40
9.67
25.07
NAMO-B
23.59
8.84
32.43
3
Navigation-only
23.59
0
23.59
NAMO
15.40
9.67
25.07
4
Navigation-only
23.59
0
23.59
TABLE I: Task-cost comparison for high-level planning in Scenes 1–4.
Test
Setting
Success (%)
Tcomp (s)
Initial heading
−40∘
96
12.3±1.7
−20∘
100
9.7±1.1
0∘
100
5.8±0.6
20∘
96
10.2±0.7
45∘
96
13.1±1.5
Friction
μ=0.3
100
6.6±0.7
TABLE II: Manipulation robustness under varying initial heading offsets and friction conditions.
Fig. 6: Interactive navigation comparison in Simulation Scene 2. (a) and (c) Robot-body pushing, (b) arm-based side pushing [ 18 ] , and (d) our discrete-contact pushing. Panels (a), (b), and (d) execute NAMO-A, while (c) executes NAMO-B.
Fig. 7: Real-robot obstacle-pushing skills. (a) and (b) demonstrate straight and side pushing, respectively. (c) shows corner pushing, where an intermediate relocation pose enables successive straight, side, and straight pushing to guide the obstacle through the corner. (d) and (e) compare straight pushing with center-only and LCR contact, respectively.
Fig. 8: Real-world integrated NAMO execution. Scene 1 and the closed-door configuration of Scene 2 contain no feasible navigation-only route; the robot sequentially relocates obstacles through two and one intermediate relocation poses, respectively, to restore connectivity and reach the goal. In the open-door configuration of Scene 2, opening the door creates a lower-cost navigation-only route, which is selected by the planner.
Peking University, Beijing 100871, China: School of Advanced Manufacturing and Robotics · Peking University, Beijing 100871, China: College of Engineering