CAREER: Scaling Up Physical Human-Robot Interaction with Generative Simulation

The ability for robots to safely and effectively make physical contact with people represents a grand challenge in robotics, with the potential to unlock transformative applications in healthcare, personal assistance, and collaborative manufacturing. However, progress has been fundamentally constrained by a critical bottleneck: the scarcity of large-scale representative data for training and evaluating robotic systems before they interact with real people. This Faculty Early Career Development Program (CAREER) award addresses that bottleneck by developing a new research paradigm called Generative Simulation for Physical Human-Robot Interaction, which will enable robots to learn from a virtually unlimited supply of automatically generated, physically realistic interaction scenarios. The work will advance the scientific foundations of safe physical human-robot interaction and accelerate the development of assistive robots that can improve quality of life and reduce the burden on caregivers. Educational activities will integrate the research platform into undergraduate and graduate robotics courses and into a residential summer program for high school students, empowering students to explore the challenges of deploying artificial intelligence in safety-critical physical systems. This CAREER project will establish the principles of generative simulation as a paradigm for developing and validating physical human-robot interaction systems. To overcome the limitations of manually designed simulations, the research will create a generative engine that systematically compiles natural language task descriptions into structured, physically plausible simulation scenarios. This pipeline will autonomously synthesize populations of virtual human partners featuring parameterized soft tissue dynamics and clinically validated biomechanical constraints. Unlike earlier work that relied on static environments, this project weaves together this massive scale of generated data with a hybrid imitation and reinforcement learning framework. By framing physical interaction as a cooperative multi-agent problem, the research will co-generate intent-driven human behaviors to train responsive and adaptive robot control policies. To strengthen safety and bridge the simulation-to-reality gap, the research will deploy an adversarial generation framework to automatically discover and mitigate policy vulnerabilities through iterative retraining. These hardened policies will then be transferred to physical robotic hardware using real-time force feedback and segmented point clouds. The goals are to produce generalizable control strategies, validate them through rigorous human participant studies alongside physical and occupational therapists, and release an open-source software suite and benchmarks. The educational plan, including mentoring doctoral students, will further accelerate these efforts by integrating generative simulation into robotics courses and leveraging student-generated adversarial scenarios to actively refine the learned robot policies, pushing the boundaries of safe, collaborative robotics. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.