Construction robots drop hand-written avoidance rules: LiDAR-only corrections cut bricklaying path irregularity by 39.2%

Construction robots can now learn avoidance corrections from unlabeled human demonstration trajectories instead of first defining social costs or annotating preferred actions. Earlier social-navigation methods depend on hand-crafted interaction rules, labeled demonstrations, or robot-specific control policies, which are hard to scale and transfer on sites where representative interaction data are limited and social responses are rarely annotated. Training data came from 74 simulated trials run by architecture, engineering and construction professionals; 36 manually screened smooth, collision-free demonstrations were used, with targets generated automatically by constant-velocity extrapolation over 28622 frames.

The correction is applied as a spatial offset to waypoints produced by an existing planner, so swapping the local planner does not require retraining. End-to-end policies that predict control commands directly are often tightly coupled to a particular robot or planner, limiting integration into existing construction-robot navigation systems. The model outputs a two-dimensional path-correction vector and was compared with original paths, social-force-model offsets and an MLP interaction-pooling baseline on the bricklaying trial and seven Dynablox sequences.

At deployment, frame-wise LiDAR detections are associated by an extended Kalman filter into continuous objects with identity and velocity, and the graph is built online from those tracks. Dynablox segments dynamic points but does not maintain persistent object identities or object-level motion states, so its output cannot directly supply the temporally consistent positions and velocities the graph model needs. In the bricklaying trial the object-detection module ran on the robot's onboard CPU at a rate corresponding to the local planner's control frequency, with no camera observations.

In the bricklaying trial the corrected path was smoother than the original while keeping the same closest pedestrian distance. The original path made no anticipatory adjustment for nearby workers and had a path irregularity of 7.22. In the bricklaying trial path irregularity fell from 7.22 to 4.39, a 39.2% reduction; path length rose from 58.96 to 63.25, and closest pedestrian distance matched the original at 1.56.

The demonstrated scope is a quadruped robot with a Velodyne VLP-16 LiDAR on an indoor bricklaying testbed, plus real-time-rate sequential processing of recorded data from seven public Dynablox sequences; the authors propose next to evaluate across more construction tasks, environmental layouts, interaction patterns and robot platforms, and to run human-centered experiments.

The authors state that real-world validation covers a limited range of construction scenarios and a single robot platform, that the constant-velocity assumption may not represent acceleration or sudden direction changes, and that evaluation relies mainly on objective metrics such as pedestrian distance and time-to-collision rather than workers' perceived comfort, motion naturalness or social acceptability.