Phase-aware 3D safety planning and recovery for collaborative robot transport under dynamic human interference

Abstract

This thesis investigates how safe and efficient collaborative robot transport can be achieved in a conveyor-to-basket pick-and-place task under dynamic human interference. The proposed system follows a planner-centered architecture that combines layered three-dimensional workspace certification, posture-aware occupancy modeling, post-grasp safety expansion, explicit recovery supervision, comfort-aware execution, and a local XY/Z command-shaping layer that acts as a safety filter. Rather than treating safety as simple stop-and-wait behavior, the thesis addresses the more difficult problem of safe recovery when a route that was previously valid becomes invalid during transport. The approach is strongly motivated by industrial practicality. Existing robot cells often rely on vendor-provided safety zones, speed limits, and controller-level supervision, but they do not usually reason at the planning level about how a task should continue when a path becomes blocked, or about the implementation cost of adding different safety and comfort features. This thesis implements those capabilities and evaluates their cost in detail in order to identify a practical operating point. In contrast, learning-heavy pipelines require extensive training, careful reward design, and more difficult debugging in exactly the part of the system where transparent safety assurances are most important. For that reason, reinforcement learning is reserved for future high-level route preference rather than being placed in the core safety loop. The final implementation was evaluated through fixed-scenario and randomized simulation benchmarks, together with a throughput-oriented service benchmark. In the main benchmark, the layered Recovery profile achieved an observed safe-success rate of 0.80 in the fixed scenarios and 0.667 in the random scenarios. No unsafe episodes were recorded for this profile in the retained benchmark, whereas unsafe episodes were observed for the Base profile. In the throughput benchmark, Recovery completed 1.70 cubes per simulated minute on average, compared with 1.37 for Base. Within the evaluated simulation, these results indicate that dynamic recertification and explicit recovery supervision can provide a favorable safety-productivity balance without reinforcement learning in the core decision loop. They do not constitute formal safety certification or a universal physical-safety guarantee.

Date	16 Aug 2026
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Michel Kadoch (Supervisor)