Follow-Bench: RPF Motion Planning Benchmark

• Follow-Bench is a unified motion planning benchmark for robot person following that standardizes evaluation of planners with metrics on safety and comfort.
• It simulates diverse scenarios including varied target trajectories, crowd dynamics, and environmental layouts to thoroughly assess planner performance.
• Re-implemented algorithms reveal trade-offs between collision avoidance and proxemic comfort, informing future research in socially-aware robotics.

Follow-Bench is a unified motion planning benchmark for robot person following (RPF), establishing a systematic simulation and evaluation framework that quantifies both safety and comfort in socially-aware robotic pursuit tasks. It provides the first end-to-end paper of RPF planners in challenging environments, re-implements and normalizes the evaluation of six major motion planning algorithms, and delivers quantitative insights into the performance trade-offs of leading approaches through both simulation and real-world experiments (Ye et al., 13 Sep 2025).

1. Definition and Objectives

Follow-Bench addresses robot person following (RPF), where mobile robots autonomously follow a designated human target in varied real-world scenarios such as personal assistance, security patrol, eldercare, and logistics. The benchmark’s primary objective is to enable rigorous comparison of RPF planners, focusing explicitly on “socially-aware” behavior—meaning the robot must continuously pursue the target while ensuring both safety (collision avoidance and reliable tracking) and comfort (adhering to proxemic norms and minimizing disturbance to bystanders).

Follow-Bench contributes a standardized testbed for end-to-end motion planning evaluation, including (i) systematic scenario design, (ii) unified metrics and logging for safety-comfort analysis, and (iii) diverse planner re-implementations within a consistent framework.

2. Scenario Design and Taxonomy

Follow-Bench simulates a spectrum of RPF scenarios along three principal axes:

• Target Trajectory Patterns: Smooth paths (L-turn, square, circular, figure-8, S-shape), sharp turns (U-turn, triangle), and special cases (“rotate then walk”). These assess the planner’s agility and tracking robustness.
• Dynamic Crowd Flows: Pedestrian counts (5–30), flows (parallel, perpendicular, circular, random), and crowd densities. These stress-test the planner’s collision avoidance and comfort management when navigation complexity increases.
• Environmental Layouts: Corridors of varying widths, doorways, intersections, cluttered spaces with scatter obstacles—parameterized to reflect typical real-world venues (mall, office, hospital).

Each scenario is constructed using parametric modules to allow controlled variation of features such as obstacle count, passage width, and flow density.

3. Motion Planning Algorithms and Re-implementation

Six widely-used RPF planners are re-implemented under the benchmark for direct comparison:

Planner Type	Core Features	Safety-Comfort Mechanisms
Social Force Model (SFM)	Goal-oriented force balancing, social cost terms	Collision avoidance, proxemics
Dynamic Window Approach (DWA)	Velocity sampling, obstacle avoidance	Extension with trajectory prediction, comfort distance
Model Predictive Control (MPC)	Optimization-based control, future trajectory prediction	Variants with dynamic search fields or target trajectory tracking

Planners are instrumented with explicit mechanisms for:

• Maintaining a comfortable following distance (proxemic compliance, typically 1.0–1.5 m).
• Recovery when target is lost (search modules, e.g., Kalman filtering).
• Avoidance of entering other pedestrians’ private zones.

Insights from simulation and hardware deployment are provided:

• MPC with trajectory prediction consistently achieves highest overall Success Rate (SR), Avoidance Success Rate (ASR), and Target Visibility Ratio (TVR).

4. Safety and Comfort Metrics

Follow-Bench introduces a set of quantitative metrics for rigorous evaluation:

Safety Metrics:

• Success Rate (SR): Percentage of runs with ≤1 collision and uninterrupted target visibility.
• Avoidance Success Rate (ASR): Fraction of runs without any collision.
• Target Visibility Ratio (TVR): Proportion of time target is visible during pursuit.
• Additional: Search Success Rate (SSR), Success weighted by Path Length (SPL) when target is lost.

Comfort Metrics:

• Path length, average velocity, acceleration, and jerk (movement smoothness).
• Time spent within or outside the target’s personal distance zone.
• Time avoided in other pedestrians’ private zones.

The proxemic distance is quantified as $|d_{rh} - R_h - R_r|$, where $d_{rh}$ is the distance between robot and target, $R_h$ the human radius, and $R_r$ the robot radius.

5. Simulation and Real-world Evaluation

Extensive simulation experiments measure each planner’s performance under controlled scenario variations. Evaluation is performed on real robot platforms (differential-drive configuration) to validate simulation observations:

• MPC variants outperform others in SR, ASR, and TVR, particularly in crowded or cluttered environments.
• Safety-comfort trade-offs are identified: strict side-following (for perceived comfort) increases risk of collision in dense crowds; back-following improves collision avoidance and tracking robustness but may compromise social engagement.

The experiments demonstrate that planners need to dynamically modulate their following strategy to balance continuous target visibility and collision minimization with socially acceptable inter-personal distances.

6. Safety-Comfort Trade-offs and Open Challenges

A key finding is the persistent safety-comfort trade-off: Proxemically comfortable following positions (side-following, close distance) lead to more frequent occlusions and collisions in dynamic, high-density scenarios. Back-following allows robust avoidance but can be less socially optimal.

Open challenges highlighted include:

• Designing planners that adaptively balance safety and comfort using hierarchical or context-aware strategies.
• Improving spatial-temporal obstacle representation beyond cluster-based polygon approaches for continuous and highly dynamic obstacles.
• Integrating advanced motion prediction (e.g., trajectory forecasting for surrounding pedestrians, deep-learning–based approaches such as Social-GAN) to enhance collision avoidance and social compliance.
• Leveraging long-term environment understanding for more robust planning (e.g., statistical modeling of typical pedestrian flows).

7. Future Research Directions

The analysis suggests several prospective directions to advance RPF algorithms:

• Decoupled hierarchical planning: sample candidate following positions optimizing visibility and comfort, then apply low-level planners for dynamic execution.
• Enhanced representation domains for obstacles and proxemic zones, especially in contextually sensitive scenarios (wide open spaces vs. corridors).
• Joint optimization over safety and comfort metrics, possibly integrating semantic environment understanding through richer simulation or real-world datasets.
• Adaptive prediction and uncertainty management for robust motion planning under real-world variability.

Conclusion

Follow-Bench represents a comprehensive and unified framework for benchmarking robot person following motion planners with dual emphasis on safety and social comfort. It systematically exposes the strengths and limitations of state-of-the-art planning methods, provides robust quantitative comparisons, and identifies substantial research opportunities for the development of adaptive, socially-aware robotic assistants in complex, human-populated environments (Ye et al., 13 Sep 2025).