SIGFPT 17Oct25: Maneuver Automata - Robotics Protocols and Dynamics

Participants: _vgr, anurajenp, ediblebadger, mtraven, zoesh, waveywaves., giovanni.merlino, maparent, timber1997

The SIGFPT group explored maneuver automata—a concept from robotics—as a formal framework for modeling protocols with natural dynamics. Participants reviewed the Frazzoli 2005 paper and related references to understand how maneuvers (non-equilibrium dynamics) could be systematically applied to protocol design. The discussion covered differential game theory's relevance to protocol sequences, particularly in scenarios like pursuit-evasion, and how negotiation dynamics could be formalized using state machines tracking proposal acceptance, conformance, and goal compatibility.

A significant theme emerged around the tension between optimization and interpretability: while deep learning can produce effective control systems, the inscrutability of such systems poses problems for human-interactive applications. Participants noted that existing AI-robot systems employ hybrid approaches rather than unconstrained learning. The group also discussed how unsupervised RL-based robotics could inform protocol theory, and emphasized the need for breaking down complex AI behaviors to reduce uncanny effects and maintain human trustworthiness.

• Maneuver automata can model protocols with natural dynamics by treating trim states as equilibria and maneuvers as non-equilibrium dynamics, with connections to game theory as a subset of negotiation dynamics.
• Differential game theory provides a mathematical framework for analyzing protocol sequences like pursuit-evasion, where specific maneuver sequences (e.g., 'double maneuver') can prove effective.
• Protocol negotiation involves states tracking proposal acceptance, conformance, goal compatibility, and the ability to withdraw unilaterally, forming a structured model of human interaction.
• There is tension between AI systems optimized through deep learning (which may be inscrutable) and the need for interpretability in systems involving human interaction to avoid uncanny or harmful effects.
• Practical AI-controlled robot systems use hybrid strategies combining LLM control with other approaches rather than pure unconstrained AI.

View discussion in Discord →