Abhinav Kumar

Robotics Department

University of Michigan

Ann Arbor

I am a Ph.D. student in Robotics at the University of Michigan at Ann Arbor, working in Professor Dmitry Berenson’s ARM Lab.

My research focuses on applying a combination of learning and planning techniques to dexterous manipulation, leveraging learning methods’ inference speed and ability to learn directly from data with the constraint satisfaction and flexibility of planning techniques. In the past, I have also done work on ensuring constraint satisfaction of machine learning models to make them more useful in robotics where hard cosntraint satisfaction and safety are important.

I also have a deep interest in public policy and national security. I have previously worked with the Special Competitive Studies Project, a think tank investigating the impacts of advanced technologies on many aspects of our lives as well as offering policy recommendations on how the government and private sector can maximize the benefit of these technologies to the American people. I have also worked with the Naval War College on studying the ethics surrounding emerging military technologies. A fun recent project was helping set up my friend’s campaign for state senate, specifically setting up and tracking the finances and designing a policy platform.

In my free time, I like to read: mostly sci-fi, fantasy, and history. I enjoy table top role playing games like Dungeons and Dragons as well as thinking about the math behind them.

Email: abhinavk99 [at] gmail [dot] com

selected publications

RA-L

Constraining Gaussian Process Implicit Surfaces for Robot Manipulation via Dataset Refinement

Abhinav Kumar, Peter Mitrano, and Dmitry Berenson

IEEE Robotics and Automation Letters, 2024

Abs arXiv HTML PDF

Model-based control faces fundamental challenges in partially-observable environments due to unmodeled obstacles. We propose an online learning and optimization method to identify and avoid unobserved obstacles online. Our method, Constraint Obeying Gaussian Implicit Surfaces (COGIS), infers contact data using a combination of visual input and state tracking, informed by predictions from a nominal dynamics model. We then fit a Gaussian process implicit surface (GPIS) to these data and refine the dataset through a novel method of enforcing constraints on the estimated surface. This allows us to design a Model Predictive Control (MPC) method that leverages the obstacle estimate to complete multiple manipulation tasks. By modeling the environment instead of attempting to directly adapt the dynamics, our method succeeds at both lowdimensional peg-in-hole tasks and high-dimensional deformable object manipulation tasks. Our method succeeds in 10/10 trials vs 1/10 for a baseline on a real-world cable manipulation task under partial observability of the environment.
Diffusion-Informed Probabilistic Contact Search for Multi-Finger Manipulation

Abhinav Kumar*, Thomas Power*, Fan Yang, and 4 more authors

2024

Abs arXiv HTML PDF

Planning contact-rich interactions for multi-finger manipulation is challenging due to the high-dimensionality and hybrid nature of dynamics. Recent advances in data-driven methods have shown promise, but are sensitive to the quality of training data. Combining learning with classical methods like trajectory optimization and search adds additional structure to the problem and domain knowledge in the form of constraints, which can lead to outperforming the data on which models are trained. We present Diffusion-Informed Probabilistic Contact Search (DIPS), which uses an A* search to plan a sequence of contact modes informed by a diffusion model. We train the diffusion model on a dataset of demonstrations consisting of contact modes and trajectories generated by a trajectory optimizer given those modes. In addition, we use a particle filter-inspired method to reason about variability in diffusion sampling arising from model error, estimating likelihoods of trajectories using a learned discriminator. We show that our method outperforms ablations that do not reason about variability and can plan contact sequences that outperform those found in training data across multiple tasks. We evaluate on simulated tabletop card sliding and screwdriver turning tasks, as well as the screwdriver task in hardware to show that our combined learning and planning approach transfers to the real world.