FocusNav: Spatial Selective Attention with Waypoint Guidance for Humanoid Local Navigation

TL;DR

This work tackles robust local navigation for humanoid robots operating in unstructured and dynamic environments. It introduces FocusNav, a spatial selective attention framework that anchors perception to predicted collision-free waypoints via Waypoint-Guided Spatial Cross-Attention (WGSCA) and dynamically modulates perceptual scope with Stability-Aware Selective Gating (SASG), trained end-to-end with a privileged GuideOracle supervisor. Key contributions include the collision-free waypoint predictor with a backward prediction paradigm and latent-consistent autoregressive decoding, a BEV-based multi-modal perception pipeline, and a dual-layer attention mechanism that balances long-range planning with immediate foothold safety; extensive experiments on the Unitree G1 show superior navigation success, reduced collisions, and enhanced motion stability in both simulated and real-world scenarios. The results demonstrate practical impact for deploying agile humanoid navigation systems in complex environments, enabling safer and more reliable operation in the presence of dynamic obstacles and uneven terrain, with potential extensions to richer semantic understanding and omnidirectional perception.

Abstract

Robust local navigation in unstructured and dynamic environments remains a significant challenge for humanoid robots, requiring a delicate balance between long-range navigation targets and immediate motion stability. In this paper, we propose FocusNav, a spatial selective attention framework that adaptively modulates the robot's perceptual field based on navigational intent and real-time stability. FocusNav features a Waypoint-Guided Spatial Cross-Attention (WGSCA) mechanism that anchors environmental feature aggregation to a sequence of predicted collision-free waypoints, ensuring task-relevant perception along the planned trajectory. To enhance robustness in complex terrains, the Stability-Aware Selective Gating (SASG) module autonomously truncates distal information when detecting instability, compelling the policy to prioritize immediate foothold safety. Extensive experiments on the Unitree G1 humanoid robot demonstrate that FocusNav significantly improves navigation success rates in challenging scenarios, outperforming baselines in both collision avoidance and motion stability, achieving robust navigation in dynamic and complex environments.

Figures (11)

• Figure 1: Snapshots of dynamic obstacle avoidance on stairs. FocusNav anchors perception to navigational intent and adaptively maintains motion stability, significantly enhancing navigational performance in dynamic and complex scenarios.
• Figure 2: Schematic of the PPA neural circuit paradigm. The system extracts environmental features through perception and predicts waypoints along the path by integrating the target point with the current location. An attention mechanism is then employed to focus on critical environmental information, enhancing mobile navigation performance.
• Figure 3: The terrain used for simulation training includes rugged terrain such as stairs, slopes, and gaps. The environment features dense static obstacles such as forest-like pillars, along with dynamic obstacles represented by red robots following randomized trajectories.
• Figure 4: Overview of the FocusNav framework. (a) Multi-modal perception encoder fuses spatially aligned LiDAR and depth camera data, utilizing voxel feature encoding and convolutional feature extraction to generate BEV feature of the environment. (b) Collision-free waypoint predictor encodes BEV feature and the robot’s proprioceptive states. Combined with the navigation goal, a goal-conditioned autoregressive decoder performs backward prediction of trajectory waypoints, enhancing the rationality of the trajectory. (c) Dual-layer spatial selective attention employs cross-attention, using predicted waypoints to guide the model’s focus toward high-relevance features along the planned trajectory. Furthermore, a Gumbel Softmax-based gating mechanism dynamically shifts the focus toward local terrain features beneath the robot's feet according to its stability.
• Figure 5: FoV of onboard LiDAR and depth camera for the robot. (a) Complementary sensing characteristics unitree_g1_2026: The LiDAR is characterized by an extended effective measurement range, whereas the depth camera reliably covers the proximal terrain surrounding the robot's feet. (b) Cross-modal point cloud in simulation: The red and blue point clouds are from the LiDAR and the depth camera, respectively.