SHIELDS / AISafety
Safe and Human Inclusive Embodied Learning for Dynamic Systems

Project Description

Official name: AIロボティクス分野等の安全性／金広・Benallegue

Learning-based motion policies enable flexible and high-performance robot behaviors, but their deployment in physical systems remains constrained by safety risks arising from distribution shift, modeling errors, time-coupled dynamics such as balance, unexpected physical contact, sim-to-real mismatch, and misinterpretation of human intent. Project SHIELDS develops a unified framework for execution-level safety assurance in AI-driven robotic motion generation. The central approach is to treat learned policies as proposals while enforcing physical safety through hybrid model-based control layers and complementary learning mechanisms that reduce risk. The project investigates optimization-based action projection using QP and control barrier functions, predictive safety enforcement using MPC, safe integration of reinforcement learning within stabilizing controllers, model-based correction of out-of-distribution behavior, failure and fall prediction with preemptive recovery, adversarial robustness training, and on-robot learning under strict safety gating. It also addresses safe physical interaction through simulation-based force modeling, safety-aware grounding of language and perception to prevent hazardous actions caused by misunderstanding, and hardware-software co-design for intrinsically safe dexterous manipulation. The project aims to deliver reusable runtime safety modules, quantitative evaluation protocols, and validated demonstrations for safe learning-enabled robotic systems.

PI in JRL

Mehdi Benallegue

Period

2026.1 - 2027.3

Funded by

BRIDGE (Bridging the gap between R&D and the IDeal society (Society 5.0) and Generating Economic and social value)