Handcrafted Safety Rules Fail to Cut Robot Collisions in Simulated Industrial Fleets

A team of robotics researchers has reported that a carefully engineered “safety-augmented” coordination scheme for robot fleets did not actually make a simulated industrial workspace safer. In a study published in Discover Robotics, Bernard Kyiewu, Clinton Amponsah, Linda Bessa-Simons, Andrew Oppong-Asante and Caleb Boakye Yiadom evaluated five coordination policies in a simulated industrial cyber-physical system built on the Multi-Agent Particle Environment, a continuous two-dimensional simulation framework. Their central and somewhat counterintuitive result: the composite Safety-Augmented Coordination Policy, or SACP, which combined proportional navigation, static workstation assignment, artificial potential-field repulsion and a handcrafted four-dimensional communication vector, recorded the highest collision frequency of any policy tested, while simpler baselines completed the task more reliably.

The experimental setting was deliberately controlled. Each policy was evaluated over 30 paired simulation episodes involving four robots and four workstations, with every episode lasting 150 timesteps. The paired design mattered: for each random seed, one valid, non-overlapping initial configuration of robots was generated and reused across all policies, so differences in outcomes could not be attributed to different starting positions. Robots were modeled as circular entities of radius 0.25 meters in a workspace spanning 1.2 meters in each direction from the origin, with Gaussian actuator noise injected into their motion. A collision was counted when the center-to-center distance between a robot pair fell below twice the robot radius, and a transition-based definition prevented a single sustained contact from being counted as multiple events.

The five policies spanned a spectrum of sophistication. Random navigation served as an uncoordinated lower bound, sampling desired velocities without any workstation information. Greedy navigation dynamically assigned each robot to its nearest unclaimed workstation. Greedy+APF added artificial potential-field repulsion, in which nearby robots generate repulsive influences that modify the final motion command. Optimal Reciprocal Collision Avoidance, an established method in which interacting agents share responsibility for selecting collision-avoiding velocities, provided a strong reactive benchmark. SACP, the proposed method, layered all of these ideas together: static index-based workstation assignment, target-directed proportional navigation, distance-dependent repulsion, and a manually specified message vector containing normalized target distance, horizontal and vertical repulsion components, and a binary workstation-arrival indicator.

The results were unambiguous on the task-performance side. Greedy navigation achieved 100 percent ever-reached workstation coverage and simultaneous full coverage in 97 percent of episodes, completing tasks faster than SACP among successful runs. Random, by contrast, reached only 9.2 percent of workstations on average and achieved simultaneous full coverage in just 3 percent of episodes. SACP landed in the middle on coverage, reaching 85 percent of workstations at least once and achieving simultaneous full coverage in 50 percent of episodes, with an intermediate mean reward of −233.32. Greedy+APF achieved an 83 percent success rate and ORCA matched Greedy at 97 percent.

On safety, however, the picture inverted. SACP recorded 16.57 collision events per episode on average, with a standard deviation of 19.37, compared with 11.57 plus or minus 14.85 for Greedy and 10.67 plus or minus 7.96 for Random. The authors are careful about what this does and does not show: because no inferential or equivalence tests were applied, overlapping distributions do not establish statistical similarity, and the comparison is limited to observed differences in central tendency and variability. Still, the available results do not show that SACP improved collision-frequency safety. The researchers attribute the pattern to concrete design issues, chiefly the static assignment rule, which directed robots across one another’s paths even when closer workstations were available, and the reactive nature of the repulsion mechanism, which responded only after robots entered the safety zone and could produce oscillation and repeated boundary crossings when several robots converged simultaneously.

The handcrafted communication layer fared no better as a safety mechanism. Every robot broadcast its four-dimensional message at every timestep, and all messages were delivered within the same timestep, with additive Gaussian noise applied on receipt. Communication activity, measured as the mean magnitude of the message vectors, varied across operational phases, rising from about 0.30 early in a representative episode to a peak of roughly 0.64 around steps 55 to 60 before declining again. But the authors stress that message magnitude alone establishes neither communication efficiency nor effectiveness. Because an otherwise identical no-communication condition was never evaluated, the study cannot determine whether the communication correction improved collision avoidance, coverage, or completion time at all. The simulation also modeled no finite bandwidth, packet loss, delay, or network contention, so no claims about channel saturation are possible.

The scalability experiments extended the evaluation of SACP across fleet sizes from two to six robots, using 20 episodes of 120 timesteps per fleet size in a fixed workspace. Raw collision events climbed steeply, from 0.65 per episode at two robots to 71.55 at six robots. After normalizing by the number of possible robot pairs and episode duration, exposure-normalized collision frequency followed a nonmonotonic pattern, rising from 0.0054 at two robots to 0.0275 at three, dipping to 0.0195 at four, then climbing to 0.0381 at five and 0.0398 at six. Mean reward declined from −81.21 at two robots to −542.33 at six, and full-coverage success fluctuated between 100, 45, 80, 40 and 50 percent across the five configurations. The authors explicitly decline to fit any quadratic, exponential or power-law scaling relationship to these five data points, noting that fleet size and spatial density increased together, so fleet size alone cannot be identified as the cause.

The study is candid about its boundaries. SACP is a fixed, handcrafted controller, not a learned multi-agent reinforcement-learning policy; no actor network, critic, value function, replay buffer or trainable parameter was used, and the scalar reward served purely as a performance measure. The environment was a simplified open two-dimensional particle world lacking warehouse aisles, static obstacles, nonholonomic motion, heterogeneous robots, human workers, localization errors and hardware constraints, so the results apply only to the simulated conditions and establish neither certified safety nor industrial deployment readiness. The evaluation also lacked a complete factorial ablation, meaning the individual and interaction effects of static assignment, repulsion and communication could not be isolated. Collision-event frequency, moreover, captures how often robots touch but not the physical intensity of contact, so the conclusions concern collision frequency rather than physical-risk severity.

What the paper ultimately delivers is a transparent negative result with a roadmap. In a normalized multi-metric comparison, ORCA earned the highest collision-safety score while maintaining high coverage and success, marking the strongest safety-task balance among the evaluated policies, while SACP recorded the lowest collision-safety score despite relatively high coverage. The authors recommend that future work isolate each SACP component under identical conditions, compare static assignment against nearest-workstation allocation, Hungarian matching and dynamic reassignment, add a no-communication control condition, and disentangle fleet-size effects from density effects by expanding the workspace as robots are added. Learning-based extensions using MADDPG or MAPPO, and eventually simulations incorporating communication delay, packet loss and sensing uncertainty, would follow before any hardware experiments. Until then, the message for multi-robot systems engineers is a sobering one: layering intuitive safety heuristics on top of one another is no substitute for established collision-avoidance mechanisms or for the controlled experiments needed to prove each layer earns its place.

Subject of Research: Simulation study of safety-aware heuristic coordination policies for multi-robot task coverage in industrial cyber-physical systems

Article Title: Safety-aware multi-robot coordination in industrial cyber-physical systems using potential fields and handcrafted communication

Article References: Kyiewu, B., Amponsah, C., Bessa-Simons, L., Oppong-Asante, A., & Yiadom, C. B. (2026). Safety-aware multi-robot coordination in industrial cyber-physical systems using potential fields and handcrafted communication. Discover Robotics, 2(1), Article 25. https://doi.org/10.1007/s44430-026-00039-z

Image Credits: AI Generated

DOI: 10.1007/s44430-026-00039-z

Keywords: multi-robot coordination, industrial cyber-physical systems, artificial potential fields, collision avoidance, handcrafted communication, simulation, task coverage, scalability, Multi-Agent Particle Environment, ORCA, heuristic control, Safety-aware