Hardik Parwana

I am a PhD candidate at the University of Michigan who just defended his PhD! I work with Prof. Dimitra Panagou as a member of the DASC lab. My research is centered on online learning of controller parameters for adaptation to novel environments under multiple state and control input constraints.

Recent News

November 2024

July 2024

I defended my PhD successfully! See a recording here.

I presented my paper "Algorithms for Finding Compatible Constraints in Receding-Horizon Control of Dynamical Systems" at ACC in Toronto

Research Highlights

Model-Predictive Framework for Online Constrained Learning of Controller Parameters under Stochastic Uncertainties

Recursive feasibility guided optimal parameter adaptation of differential convex optimization policies for safety-critical systems,

IEEE International Conference on Robotics and Automation (ICRA) 2022

Foresee: Model-based reinforcement learning using unscented transform with application to tuning of control barrier functions

(under review)

How can we adapt any parametric controller for better performance while guaranteeing safety?

In this work, I demonstrate the application of a constrained gradient descent scheme to learn parameters in a model predictive framework. The quadrotor, shown in the image below, automatically adapts its controller to better follow the desired reference while avoiding the obstacle.

When the known dynamics are uncertain at best, how can we predict future state distributions (for use in a model-predictive framework) with a finite number of weighted samples?

In this work, I use an unscented transform and propose differentiable Expansion Compression layers for scalable uncertainty propagation under state-dependent disturbances. The figure below compares how Monte Carlo (blue) and Unscented Transform particles (yellow and orange for two variations of UT) evolve with time for an artificially designed dynamics model. Yellow only tries to preserve mean and covariance whereas orange also preserves skewness (for asymmetry) and kurtosis.

Trust-based tuning of Control Barrier Functions for Uncooperative Multi-Agent Systems

Trust-based rate-tunable control barrier functions for non-cooperative multi-agent systems.

IEEE Conference on Decision and Control (CDC) 2022

Rate-tunable control barrier functions: Methods and algorithms for online adaptation.

(under review)

How can robot controllers adapt their aggressiveness to apriori unknown agents with varying degrees of cooperativeness, that is, agents whose motives range from being cooperative, or non-cooperative, to being adversarial?

In this work, I introduce a trust factor to quantify other agents' contribution to safety on a continuous normalized scale of 0 to 1. Then, I define parameter dynamics as dependent on the trust factor to change the controller parameters and generate customized responses for each agent in the environment.

Compatibility of Control Barrier Functions under Multiple Barrier Constraints

Recursive feasibility guided optimal parameter adaptation of differential convex optimization policies for safety-critical systems

IEEE International Conference on Robotics and Automation (ICRA) 2022

Rate-tunable control barrier functions: Methods and algorithms for online adaptation.

(under review)

Feasible Space Monitoring for Multiple Control Barrier Functions with application to Large Scale Indoor Navigation

(under review)

How do we ensure the compatibility of a controller that imposes multiple barrier constraints?

See our papers for multiple methods to improve the feasibility of CBF controllers. Our two approaches include - 1) adapting CBF's classK function parameters for better feasibility (and performance), 2) monitoring the size (volume) of feasible space of the QP controller and ensuring its infeasibility by preventing the feasible space volume from going to zero.

Constraint Removal under Additive Prioritization

Algorithms for Finding Compatible Constraints in Receding-Horizon Control of Dynamical Systems

IEEE American Control Conference (ACC) 2024

Which constraints should a controller (framed as an optimization problem) disregard when it is impossible to satisfy all?

Under additive priorities, for example, when minimum constraints must be removed, this problem is known to be NP-Hard. We propose a heuristic, the first of its kind to address this problem in the context of dynamical systems, that exploits information from feasible controller instances in the past to make an informed prior (guess) on constraints to be removed when the controller becomes infeasible for the first time.

Control barrier function based controller for navigation amonsgt humans in narrow spaces

Development and Evaluation of CBF and MPPI Control Algorithms for Social Navigation

See my papers on Model Predictive Path Integral (MPPI) Controller and Control Barrier Function (CBF) Controllers for navigation in human-occupied constrained spaces like hospitals.

Feasible Space Monitoring for Multiple Control Barrier Functions with application to Large Scale Indoor Navigation

(under review) arXiv:2312.07803

Risk-aware MPPI for Stochastic Hybrid Systems

(under review) arXiv:2411.09198 (2024)

Control barrier function based controller for navigation amonsgt humans in narrow spaces

Model Predictive Path integral (MPPI) Controller for navigation with stochastic hybrid human dynamics model. Humans switch dynamics when the robot is in their sensing region. See my paper for more results like this and for comparison with Risk-Aware MPPI.