ExoSLM: a domain-specific small language model for requirements-to-parameter translation in simulation-based exoskeleton assistance design

Exoskeletons allow unprecedented human mobility, but the laws of controls and a high manual design effort restrict areal motions. To ease such constraints in assist-profile development, we introduce ExoSLM, a domain-specific small language model (SLM), as a two-stage design aid, from requirements to parameters. ExoSLM offers natural-language encoding and decoding for explicit conditioning in biomechanical simulation states. This methodology enables high-level text to naturally express functional evaluation specifications and at the same time to produce a structured, simulation-executable parameterization for agent scheduling. This provides the convenience of using basic text as the user interface for entering evaluation criteria while also enabling direct mapping of that text into the design and implementation of control theory concepts for a user-in-the-loop control. Key to this utility as an assistance design tool, the model counts the number of tokens k from the MARKER token it emits until the terminal token is emitted, and assumes that the first k tokens emitted from the conditioned output encode the value k for the current sequence.