Data-driven safe iterative control

01 · Learning from successful trajectories

Stochastic optimal control for iterative tasks

Safe Information-Theoretic Learning Model Predictive Control (SIT-LMPC) addresses a constrained infinite-horizon optimal control problem for discrete-time nonlinear stochastic systems. Trajectories from previous task executions are used to learn a value function with normalizing flows.

To solve the resulting stochastic receding-horizon optimal control problem, we use AP-MPPI, am information-theoretic model predictive control (MPC) framework that balances safety and optimality by online tuning of penalty parameters to enforce the constraints. The algorithm is designed for parallel execution on graphics processing units. Simulations and hardware experiments demonstrate iterative performance improvement while satisfying system constraints.

Hardware results: lap time and boundary violations

02 · Learning from failed executions

State-control invariance from failure data

Failure-Aware Iterative Learning (FAIL) computes the maximal state-control invariant set for deterministic linear time-invariant systems with unknown dynamics and polytopic constraints. This set encodes both the maximal control invariant state set and, at each state, the control inputs that preserve invariance.

FAIL uses regression on failing trajectories to learn predecessor halfspaces from one-step failing state-input pairs. These halfspaces update the constraints in the joint state-control space. We prove that the learned constraint set converges monotonically to the maximal state-control invariant set without knowledge of the system dynamics.

03 · Ongoing research

Reducing the number of failures

Learning from failures raises a further question: how many failed executions are required to identify the maximal state-control invariant set? My ongoing research investigates how to select informative executions to reduce this number.

A longer-term objective is to combine this constraint-learning approach with performance improvement from successful trajectories. Such an integration would allow a controller to use both types of execution data within a common iterative control framework.

04 · Autonomous vehicle applications

Online model learning and adaptive control

In autonomous vehicle applications, I study how online system identification can account for changing friction, nonplanar terrain, and model uncertainty. This work combines Gaussian process dynamics models with predictive control and includes ensemble models for driving across surfaces with different friction, recursive sparse models for nonplanar MPC, and adaptive differentiable control for racing.

I also develop control software and robotic platforms for evaluating these methods in simulation and on hardware, including the Roboracer R5 off-road platform.

Nonplanar MPC · Ensemble Gaussian processes · Adaptive differentiable MPCC