Modular Legged Robots

Modularity is usually sold as a free virtue: swap a payload, change a foot, link two bipeds into a quadruped. My current research — the core of my PhD at FORTH — treats it instead as generality purchased at a cost, and asks the question the field has debated for thirty years without measuring: what is that cost, in concrete engineering units?

Approach

The work is built around Modbot, a compact 8-DoF biped module designed to operate alone or linked with a partner into a quadruped:

• Design: the module family is a controlled ladder of configurations — feet, payloads, and linked bodies — so that each added degree of modular freedom can be measured, not argued about.
• Control: configuration-conditioned reinforcement-learning policies are trained at scale in GPU simulation. A single conditioned controller serves the whole family; a proprioceptive student can even infer the current configuration without being told.
• Measurement: every configuration is scored on the same ledger — reliability, control performance, and energy (cost of transport) — to map the generality–cost frontier: which modular axes pay for themselves, and which are dead weight.

Why it matters

Legged locomotion learning is now mature; what the field lacks is an honest account of whether reconfigurable machines deliver more useful field-hours than purpose-built ones. Within the RAICAM network — mobile robots for inspection and maintenance of critical infrastructure — that question is practical, not aesthetic: one adaptable platform versus five specialized ones is a procurement decision.

Status

Modbot is designed, modeled, and training in simulation; a manuscript on the generality–cost frontier and the biped-linkage benchmark is in preparation (2026). This theme also connects back to my earlier hexapod work, where topology switching was already part of the capability model.