Ship Collision Avoidance for Human–Machine Cooperation (HMI-CAS)

established Updated: 2026-05-31
collision-avoidance human-machine-interaction mass shared-decision-making velocity-obstacle colregs remote-control automation-transparency

Ship Collision Avoidance for Human–Machine Cooperation (HMI-CAS)

Status: established
Last updated: 2026-05-31
Sources: 1 S2.0 S0029801820308738 Main.Pdf
Tags: [collision-avoidance, human-machine-interaction, mass, shared-decision-making, velocity-obstacle, colregs, remote-control, automation-transparency]

Summary

Huang, Chen, Negenborn and van Gelder (2020) propose a Human-Machine Interaction oriented Collision Avoidance System (HMI-CAS) for Maritime Autonomous Surface Ships (MASS). The system makes the machine's collision-avoidance decision interpretable and interactive: it visualises the solution space, marks dangerous manoeuvres, presents an optimal solution, and lets a human operator read, judge, authorise, or override it. It is built on a generalised velocity obstacle (GVO) algorithm extended to under-actuated ship dynamics, and is validated in two- and multi-ship simulations. The authors frame human and machine intelligence as complementary rather than adversarial, with the human supplying rule interpretation and the machine supplying computation.

Body

Context

Huang et al. (2020) examine how a collision avoidance system can combine human and machine intelligence during conflict resolution, rather than assigning the task entirely to one or the other. Their method is a control-theoretic framework (GVO over under-actuated ship dynamics) plus an interface design, demonstrated in simulation. Within this knowledge base the paper supplies the technical, decision-level counterpart to the human-factors strands: it operationalises the automation-transparency and intervention concerns of Human In The Loop Automation Transparency, gives a concrete shared-control mechanism for the remote operator of Remote Operation Centres Mass, and addresses the trust-calibration problem of Trust In Human Autonomy Teaming at the level of a single navigational decision. It also names the same one-to-many supervision and SA-building problems treated in Multi Ship Remote Operations Workload Sa as open work.

Key Points

The paper starts from a gap. Existing collision avoidance systems (CASs) split into two families that do not meet: alert systems that warn manned-ship operators of danger but leave solution-finding to the human, and automatic systems that resolve conflicts for unmanned ships without an interface for human intervention (PDF p. 1, orig. p. 1). Huang et al. argue neither is sufficient, because teaching automation to understand COLREGs remains unsolved and public trust in fully autonomous ships is still questionable, while human and machine intelligence are complementary — the human is good at interpreting regulations, the machine at computation (PDF p. 1, orig. p. 1). They also note that fully autonomous and manned ships will coexist, so cooperation between the two is required in practice (PDF p. 1, orig. p. 1).

The authors classify human–machine interaction by the form of the solution the machine offers and the operations the human can then perform. The machine can deliver five solution forms: one feasible solution (u), one optimal solution (u*), a finite set of safe solutions (U), a closed collision-free region, and all closed dangerous regions (K) (PDF p. 3, orig. p. 3). Against these the human can switch to manual mode, accept the solution, modify the utility function, pick one of several offered solutions, freely choose within a feasible region, or validate an arbitrary solution of their own against the dangerous set (PDF p. 3, orig. p. 3). A survey of 14 representative algorithms across five method groups (rule-based, virtual field, discrete inputs, continuous inputs, re-planning) shows that only velocity obstacle (VO) methods can produce all five information types and therefore support the full range of interactions (PDF pp. 3–4, orig. pp. 3–4).

The proposed HMI-CAS adds an interaction channel between the machine's Guidance system and the human, layered onto a standard Guidance–Navigation–Control (GNC) architecture (PDF pp. 4–5, orig. pp. 4–5). The Guidance system holds three modules: a Global Planner (waypoint/path planning, or human-input path), a Local Planner (trajectory prediction, conflict detection, conflict resolution), and an Interface that presents solutions to the human and feeds the human's chosen solution back to the machine (PDF p. 5, orig. p. 5). The interface visualises the resolution space and the machine's selected solution so the operator can read the machine's intent, authorise it, intervene, or input a new solution; the system runs in either an autonomous (Auto) mode or a manual mode, switching to manual the moment human input is detected (PDF pp. 5–6, orig. pp. 5–6). The stated aim is that, by seeing the decision process, the operator can build trust in the system and need not stay aboard — interaction can occur from an offshore control centre (PDF p. 6, orig. p. 6).

The technical contribution is extending VO to under-actuated ships. A ship is under-actuated — thrust controls surge and torque controls yaw, but sway is not directly controllable — so the authors apply a Generalized Velocity Obstacle (GVO) algorithm with a PD high-level controller and successive linearisation (Runge-Kutta integration plus Taylor expansion) to compute the set of velocity changes that lead to collision, the UO set (PDF pp. 7–9, orig. pp. 7–9). Two control parameterisations are compared: Control Group 1 (CG1) tracks heading and surge speed; CG2 tracks course and resultant speed (PDF p. 8, orig. p. 8). In Auto mode the system ignores COLREGs and solves a quadratic optimisation that minimises change in control, weighted to favour course change over speed change as is common at sea, leaving rule compliance to the human (PDF pp. 9–10, orig. pp. 9–10).

Simulations of two-ship and multi-ship encounters with the CyberShip II model show all settings (fully actuated Standard Group, CG1, CG2) avoid collision automatically while keeping relative distances above the safety distance, and support human intervention via the visualised solution space (PDF pp. 10–15, orig. pp. 10–15). A central finding is that under-actuation causes oscillation in the avoidance manoeuvre, more severe under CG1 (no sway control) than CG2; the authors trace this to linearisation error rather than the inter-ship incoordination reported for VO by Van Den Berg et al. (2008), and recommend CG2 to reduce it (PDF p. 15, orig. p. 15).

In discussion the authors position the human and machine as mutually supporting. The machine visualises danger so the operator can judge safety, see whether the optimal solution violates rules, and validate alternatives; the human in turn keeps the system rule-compliant (drawing on seamanship that differs between navigators) and can find a collision-free solution in extreme cases where Auto mode, using a convex approximation of the feasible space, finds none (PDF p. 15, orig. p. 15). They map the current arrangement — operator must monitor at all times — onto SAE Level 2 driving automation, and identify the transition of control authority (timing, early alerting to build situation awareness, mixed control during handover) as the work needed to reach higher autonomy levels (PDF p. 16, orig. p. 16, citing SAE International, 2016).

Conclusion

Huang et al. (2020) conclude that automation in MASS need not exclude the human but should incorporate human intelligence for a safer, smarter system. Their HMI-CAS makes the machine's collision-avoidance reasoning transparent — visualising the solution space and dangerous manoeuvres — so the operator can read, understand, and intervene, while the operator supplies the rule interpretation the automation cannot yet guarantee. The velocity obstacle family is selected because it alone exposes all five solution forms needed for the full interaction set, and GVO extends it to realistic under-actuated dynamics, with CG2 recommended to limit oscillation. The authors leave the transition of control authority, physical-experiment validation, and the behavioural characteristics of maritime HMI as future work — the same handover and SA-building problems that the human-factors articles in this knowledge base treat from the operator's side.

References

Huang, Y., Chen, L., Negenborn, R.R. & van Gelder, P.H.A.J.M. (2020) 'A ship collision avoidance system for human-machine cooperation during collision avoidance', Ocean Engineering, 217, 107913. doi: 10.1016/j.oceaneng.2020.107913. huang2020collision

SAE International (2016) Levels of driving automation are defined in new SAE International standard J3016. Warrendale, PA: SAE International. To be validated.

Van Den Berg, J., Lin, M. & Manocha, D. (2008) 'Reciprocal velocity obstacles for real-time multi-agent navigation', in 2008 IEEE International Conference on Robotics and Automation. Pasadena, CA: IEEE, pp. 1928–1935. doi: 10.1109/ROBOT.2008.4543489. To be validated.

Open Questions