Organizations: Yanmei Jiao, Anpeng Lu and Wenhan Hu are with the School of Information Science and Engineering, Hangzhou Normal University, Hangzhou, P.R. China. · Rong Xiong and Yue Wang are with the State Key Laboratory of Industrial Control and Technology, Zhejiang University, Hangzhou, P.R. China. · Huajin Tang is with Neuromorphic Computing and Robotic Cognition Lab, Zhejiang University, Hangzhou, P.R. China. · Wen-an Zhang is with the Department of Automation, Zhejiang University of Technology, Hangzhou 310023, China.
Object-goal visual navigation requires robots to reason over semantic structure and act effectively under partial observability. Recent approaches based on object-level topological maps enable long-horizon navigation without dense geometric reconstruction, but their execution remains limited by the gap between global topological guidance and local perception-driven control. In particular, local decisions are made solely from the current egocentric observation, without access to information beyond the robot's field of view. As a result, the robot may persist along its current heading even when initially oriented away from the goal, moving toward directions that do not decrease the global topological distance. In this work, we propose IntentReact, an intent-conditioned object-centric navigation framework that introduces a compact interface between global topological planning and reactive object-centric control. Our approach encodes global topological guidance as a low-dimensional directional signal, termed intent, which conditions a learned waypoint prediction policy to bias navigation toward topologically consistent progression. This design enables the robot to promptly reorient when local observations are misleading, guiding motion toward directions that decrease global topological distance while preserving the reactivity and robustness of object-centric control. We evaluate the proposed framework through extensive experiments, demonstrating improved navigation success and execution quality compared to prior object-centric navigation methods.
Figures & tables
Fig. 1: Existing methods exhibit short-sighted behavior due to the decoupling of global planning and local control, leading to locally reasonable yet globally inefficient or even incorrect trajectories: (a) excessive subgoal attraction results in zigzag trajectories and low overall efficiency; (b) large initial heading errors cause persistent wrong-direction motion. Our method introduces a low-dimensional intent to condition the waypoint policy, enforcing topologically consistent progression and enabling efficient correction, as shown in (c).
Fig. 2: IntentReact Navigation pipeline. (1) We construct an object-level topological map with 3D node information. (2) During online execution, the current observation is matched to the map to obtain a set of query nodes. (3) For each query node, global planning is performed to obtain a shortest path, from which an object-level costmap is built and an intent vector indicating the direction of decreasing topological distance is computed and fed into the local control stage. (4) The intent is incorporated as a conditional signal to modulate the controller’s feature representation, enabling global guidance rather than hard intervention. (5) A BEV costmap is further used to correct the feasibility of the predicted waypoint, forming a unified closed loop between learned control and geometric feasibility.
Fig. 3: Comparison of different conditioning strategies. (a) Global path conditioning leverages full map or planned path information, but introduces redundancy and weakens local reactivity. (b) Goal-direction conditioning provides a simple directional cue, but may lead to premature turning and local collisions. (c) The proposed 2-hop intent conditioning encodes a compact topological signal, effectively balancing global consistency and local feasibility.
Method
Imitate
Alt Goal
Shortcut
Reverse
SR
SPL
SSPL
SR
SPL
SSPL
SR
SPL
SSPL
SR
SPL
SSPL
RoboHop [ 11 ]
54.63
53.13
67.72
21.30
20.63
35.32
27.62
25.98
45.12
38.89
34.89
49.82
TANGO [ 20 ]
64.81
59.21
72.67
24.07
20.57
36.85
34.29
30.56
48.90
55.56
48.48
62.28
ObjectReact-ft [ 10 ]
63.89
57.84
76.05
26.85
23.79
38.50
39.05
34.55
57.28
51.85
48.04
64.94
(Ours) IntentReact
81.48
76.12
87.85
34.26
31.07
43.69
60.95
53.86
74.35
62.04
59.89
70.18
TABLE I: Performance comparison under the GT-Topological setting.
Fig. 4: Qualitative results under the Shortcut setting. When the initial heading deviates from the goal direction, RoboHop, TANGO, and ObjectReact often follow misleading local cues, resulting in inefficient or failed trajectories. In contrast, our method leverages the intent signal to correct early deviations and maintain topologically consistent progression. In cases without ambiguity, it behaves identically to ObjectReact, indicating that intent acts as a soft bias without disrupting the underlying reactive policy.
Fig. 5: Step-by-step visualization of our method for the first case in Fig. 4 . At each step, four panels show the RGB observation, segmentation, BEV waypoint, and intent direction. Despite a large initial heading error, intent provides consistent directional bias, enabling rapid correction toward the goal (step 24). BEV refinement further ensures safe execution by projecting raw waypoints onto feasible regions (step 75).
Fig. 6: Performance under different initial orientation offsets in the Opposite task. Annotated percentages indicate the relative improvement over the strongest baseline.
Method
SR (%)
SPL
SSPL
Retention
(Base) ObjectReact-ft
72.22
66.54
83.45
–
(Ours) IntentReact ( 0∘ )
81.48
76.12
87.85
1.00
(Ours) IntentReact ( 5∘ )
82.41
77.23
88.11
1.01
(Ours) IntentReact ( 10∘ )
81.48
75.70
88.08
0.99
(Ours) IntentReact ( 20∘ )
79.63
75.24
87.14
0.99
(Ours) IntentReact ( 30∘ )
77.78
72.42
85.98
0.95
TABLE II: Sensitivity to intent estimation errors.
Method
Steps
SR (%)
SPL
SSPL
(Base) ObjectReact-ft
106.46
63.89
57.84
76.05
(Ours) Base + Intent
102.79
72.22
66.54
83.45
(Ours) Base + Intent (sign)
101.08
70.37
66.16
81.48
(Ours) Base + Intent (concat)
101.89
70.37
65.85
82.60
(Ours) Base + Intent (+dist)
103.20
70.37
65.42
81.55
(Ours) Base + BEV
104.38
75.00
68.45
80.43
TABLE III: Ablation study of the proposed components.
Method
Imitate
Alt Goal
Shortcut
Reverse
SR
SPL
SSPL
SR
SPL
SSPL
SR
SPL
SSPL
SR
SPL
SSPL
GT PR
RoboHop
48.15
43.70
55.90
19.44
14.90
27.23
17.14
13.74
27.10
12.96
10.48
19.84
TANGO
48.15
41.74
52.31
24.07
18.17
29.66
22.86
17.40
34.03
31.48
26.40
38.06
ObjectReact-ft
37.04
32.01
43.37
13.89
11.54
22.67
4.76
4.32
22.48
19.44
15.97
31.26
(Ours) IntentReact
64.81
60.49
68.46
28.70
23.56
37.36
39.05
32.83
51.13
35.19
28.16
45.98
NetVLAD
(Ours) IntentReact
65.74
58.86
67.71
25.93
20.82
37.34
32.38
27.54
48.79
39.81
34.44
51.31
TABLE IV: Navigation performance under the learned-perception setting.
Object navigation requires a robot to search for an unobserved target in an unknown environment by deciding where to explore next under partial observability. Effective search resembles human-like exploration: selectively probing visually promising frontiers while relying on spatial memory to avoid redundant revisits. We propose IntentNav, a spatial-visual imitation framework that learns human-like ObjectNav policies from human demonstrations. To infer high-level search intent from low-level human actions, we introduce Frontier-based Human-Intent Labeling, which looks ahead in human demonstrations and labels the frontier that best explains the demonstrator's future search direction. We construct a spatial-visual candidate space, where BEV memory tracks explored regions, unexplored frontiers, and trajectory history, while egocentric visual memory provides semantic cues for each candidate. A VLM policy is trained to select among these grounded candidates, using Intent-Aligned Objective to encourage consistent and human-like exploration. IntentNav achieves state-of-the-art performance on the MP3D, HM3D-v1 and HM3D-v2 ObjectNav benchmarks. The proposed candidate-level navigation interface transfers zero-shot to wheeled, quadruped, and humanoid robots without further VLM fine-tuning. Project page.
Yuxin Cai, Zongtai Li, Maonan Wang +9
Nanyang Technological University · Carnegie Mellon University · The Chinese University of Hong Kong +1
Vision-language navigation requires embodied agents to navigate environments using natural language instructions and visual observations. Existing approaches typically decompose navigation into sequential language-guided decisions or rely on online exploration without prior environmental knowledge. Scene graph representations offer compact semantic memory but remain decoupled from downstream navigation, which still depends on dense metric maps. To close this gap, we propose an object--path graph that unifies open-vocabulary semantic reasoning with topological navigation. The proposed representation jointly supports semantic grounding, graph-based localization, and navigation within a single lightweight topological framework. Building on this graph, we introduce a navigation strategy that combines global path planning with local inter-node execution through lightweight node localization and semantic visual servoing, enabling navigation directly over the graph without dense metric reconstruction. Experiments on HM3D and Replica demonstrate competitive performance in open-vocabulary object grounding through the proposed hierarchical graph structure, while achieving effective navigation performance. Real-world robot experiments further validate the practicality of the proposed framework.
Linwei Zheng, Daojie Peng, Bingtao Wang +2
The Hong Kong University of Science and Technology (Guangzhou) · The Hong Kong University of Science and Technology
Multi-goal robotic tasks are commonly delegated to planning, because reactive control is prone to local minima when objectives conflict. We show that many such failures stem from static goal representations, not task complexity: the designer defines the objectives, but where and when they must be traded off follows from the world's interaction structure which changes with the state. To read this structure, we extend Active InterCONnect (AICON), a graph of recursive estimators coupled by differentiable interconnections, with adaptive nullspace projections: lower-priority gradients enter the nullspace of higher-priority ones wherever they meet along the graph, not only in action space, ordered by current gradient magnitudes, not a fixed hierarchy. Where two gradients oppose and no weighting of them makes progress, the system instead explores in the nullspace of the stronger one until the conflict dissolves. We robustly solve 100 non-convex navigation and 100 pushT problems, outperforming static potential fields, a diffusion policy, and the same method without projection. Unmodified but embedded in a larger graph, it absorbs perceptual uncertainty, joint limits, and self-collisions on a real robot, solving 49 of 50 pushing trials across varied objects, with active camera control and disturbance recovery emerging from the same coupling. Much of the behavior attributed to planning may thus be within reach of control, once goal representations compose from the conflicts they encounter.
Vito Mengers, Oliver Brock
Robotics and Biology Laboratory, Technische Universität Berlin · Science of Intelligence (SCIoI), Cluster of Excellence, Berlin, Germany · Robotics Institute Germany