I am a research scientist at the Toyota Research Institute, where I develop decision-making algorithms for autonomous systems. My research leverages tools from optimal and stochastic control, differential geometry, and machine learning, revealing insights for designing fast and reliable methods with optimality, accuracy, and adaptation guarantees. It was deployed on drones, spacecraft and Mars rover testbeds, rockets, and mobile robotic manipulators.
Some problems I have been working on:
- Trajectory optimization under uncertainty: How can we compute optimal trajectories for uncertain nonlinear dynamical systems that explicitly account for the risk of failure?
- Forward reachability analysis: How can we propagate uncertainty through complex systems (potentially with neural networks in the loop) in milliseconds and use it for planning and control? What accuracy can we expect? What problem structure can we exploit?
- Reliable learning-based control: How can a system safely (meta-)learn its dynamics, while handling the exploration-exploitation tradeoff and always satisfying constraints?
- Resilient navigation: How can a drone autonomously fly even if all its exteroceptive sensors have failed?
- Vision-based control: How can a robot achieve precise control from high-dimensional visual inputs while ensuring safe and zero-shot deployment?
Project Page / Paper / Code - We study the problem of estimating the convex hull of the image of a compact set with smooth boundary.
Project Page / Paper / Code - We revisit the sample average approximation (SAA) approach for general non-convex stochastic programming and apply the method to stochastic optimal control problems.
Project Page / Paper / Code - We give a finite-dimensional characterization of the convex hulls of reachable sets of nonlinear systems (dx/dt=f(x)+w).
Project Page / Paper / Code - A sample-based approach to risk-averse trajectory optimization.
Project Page / Paper / Blog Post / Video - We propose an effective strategy for table wiping combining the strengths of reinforcement learning and whole-body trajectory optimization.
Project Page / Paper / Code - We propose a sequential convex programming framework for non-linear finite-dimensional stochastic optimal control.
Project Page / Paper / Code - We analyze general SCP procedures for continuous-time optimal control.
Project Page / Paper / Video / More Hardware Results - How can robots safely learn their dynamics?
Project Page / Paper / Video / Code - New efficient sampling-based reachability analysis algorithms.
Project Page / Paper / Code - We propose an algorithm for chance-constrained trajectory optimization.
Project Page / Paper / Code - We propose a learning-based strategy to warm-start trajectory optimization algorithms.
Project Page / Paper / Video - How can a drone fly blindly, when all its exteroceptive sensors have failed?
Project Page / Paper / Launch Video / ARIS website - We propose a control algorithm for a rocket to accurately stop at a target apogee.
Project Page / Paper - We identify limitations of wheel-soil interaction models and present a new method.
Project Page / Paper / Code - A new SCP trajectory optimization algorithm on manifolds.