TLDR: Team Xiaomi EV-AD VLA developed a “Proactive Risk Perception Module” that augments the Falcon framework, enabling autonomous robots to better assess and avoid collision risks with humans in dynamic indoor environments. This module, which predicts continuous distance-based risk scores, helped their system achieve second place in the IROS 2025 RoboSense Challenge Social Navigation Track by improving personal space compliance and collision avoidance.
In the rapidly evolving world of robotics, enabling autonomous agents to navigate safely and efficiently in environments shared with humans is a significant challenge. This field, known as Social Navigation, requires robots not only to avoid collisions but also to adhere to social norms, such as maintaining appropriate personal distances. A recent technical report from Team Xiaomi EV-AD VLA details their innovative approach to this problem, which secured them second place in the IROS 2025 RoboSense Challenge Social Navigation Track.
The Challenge of Social Navigation
Traditional robot navigation often relies on static maps, but social navigation demands continuous adaptation to dynamic human behaviors. Robots must predict human movements, avoid obstructing paths, and maintain socially acceptable distances, all while pursuing their own goals. This is particularly complex in indoor environments, which often have constrained spaces and higher probabilities of close encounters.
Many existing methods, particularly those based on reinforcement learning, sometimes struggle with foresight, focusing on immediate obstacle avoidance rather than anticipating future scenarios. While some approaches integrate human trajectory forecasting, these are often designed for outdoor autonomous driving and don’t fully address the unique challenges of indoor settings.
Building on the Falcon Framework
Team Xiaomi EV-AD VLA’s solution builds upon an existing framework called Falcon, which already incorporates “future-aware” mechanisms like a Social Cognition Penalty and a Spatial-Temporal Precognition Module to improve trajectory prediction. The IROS 2025 challenge specifically required participants to develop navigation policies using only onboard sensors, including RGB-D observations (color and depth information) and odometry, without access to global maps or privileged information.
Introducing Proactive Risk Perception
The core innovation introduced by Team Xiaomi EV-AD VLA is the “Proactive Risk Perception Module.” This module significantly enhances the agent’s ability to assess and respond to potential collision risks in human-populated environments. Unlike methods that only penalize violations after they occur, this module learns to predict continuous, distance-based collision risk scores for surrounding humans. This means the robot can understand varying levels of danger, from safe zones to immediate threats, and react proactively.
The module uses a lightweight neural network that processes the robot’s internal state to generate risk scores for each nearby human. These scores are based on defined risk zones: a danger zone (e.g., within 2.0 meters), a warning zone (e.g., between 2.0 and 4.0 meters), and a safe zone (beyond 4.0 meters). By incorporating a loss function that emphasizes high-risk scenarios during training, the robot develops more nuanced spatial reasoning and anticipatory collision avoidance strategies, learning to maintain comfortable margins around people rather than just avoiding direct contact.
This risk perception module operates in parallel with Falcon’s existing auxiliary tasks (like estimating human population, position, and forecasting trajectories). It shares the same internal representation, ensuring efficient multi-task learning without adding significant computational overhead during operation. The learned risk awareness is implicitly integrated into the main navigation policy.
Also Read:
- Advancing Humanoid AI: Team Revontuli’s Winning World Models
- TrackVLA++: Advancing Embodied Visual Tracking with Spatial Reasoning and Memory
Evaluation and Results
The team’s approach was evaluated on the Social-HM3D benchmark, a large-scale, photo-realistic dataset designed for social navigation in dynamic indoor environments. This dataset features diverse scenes, goal-oriented human trajectories, naturalistic motion dynamics, and calibrated human population densities, making it a rigorous testbed.
Performance was assessed using metrics such as Success Rate (SR), Success weighted by Path Length (SPL), Personal Space Compliance (PSC), and Human Collision Rate (H-Coll). Team Xiaomi EV-AD VLA achieved an impressive second place among 16 participating teams, with a total score of 0.6994. Their method demonstrated strong goal-reaching capabilities, efficient path planning, and excellent adherence to social distancing norms, with a relatively low human collision rate given the challenge’s difficulty.
The competitive results underscore the effectiveness of integrating explicit risk quantification with trajectory prediction for improved social navigation. This work highlights a promising direction for developing autonomous agents that can safely and compliantly navigate complex human environments. For more technical details, you can refer to the full report: Team Xiaomi EV-AD VLA: Learning to Navigate Socially Through Proactive Risk Perception.
