Ashwin Balakrishna

I am a Research Scientist at Google DeepMind working on building generally intelligent robots. Broadly speaking, I am excited about bridging the general-purpose capabilities of vision and language foundation models with algorithmic paradigms for learning from interaction, such as imitation learning and reinforcement learning. I've previously spent time at Toyota Research Institute and Nuro, where I worked on large-scale robot learning for robotic manipulation and autonomous vehicle planning respectively. I did my PhD in Computer Science in the AUTOLAB at UC Berkeley and completed my bachelors degree at Caltech in Electrical Engineering. Below are selected publications which are representative of my current research interests -- please see my CV for a full account of my prior work.

Email  /  Resume  /  CV  /  Google Scholar  /  Twitter  /  LinkedIn  

Using Gemini to achieve highly generalizable and dexterous robot manipulation. My specific contributions were focused on post-training algorithms and evaluations.

A framework and benchmark to evaluate the generalization capability of robotic manipulation policies across a number of axes, including visual, semantic, and behavioral generalization.

A simple and general method to filter robotic manipulation datasets for high quality demonstrations.

A simple and general method to filter robotic manipulation datasets for high quality demonstrations.

Learning performant hierarchical imitation learning policies by improving the interface between goal proposal networks and low-level control policies.

Fully open-source vision-language-action model for general-purpose robotic manipulation. Exhibits strong performance both in the zero-shot and finetuning regimes.

Develops a safe reinforcement learning system for autonomous motion selection as part of the NuroDriver. Tested extensively in simulation and real world autonomous trials!

A large-scale cross-embodiment dataset and results suggesting positive policy transfer across robot embodiments.

A large-scale, in-the-wild robot manipulation dataset of 75K+ trajectories collected in various settings such as homes, labs, offices and more across multiple continents! Initial experiments suggest that co-training policies with DROID significantly improves policy robustness and OOD generalization.

A thorough investigation of what design choices matter most for building performant visually-conditioned language models. We release optimized and hackable training code, evaluation code, and all trained models for the community to build on.

A simple reinforcement learning algorithm that can be applied to any existing actor critic method to accelerate exploration for sparse reward tasks.

An algorithm for robust off policy comparison of learned reward functions.

Safe exploration by meta-learning risk measures across environments with different dynamics.

An algorithm for query-efficient interactive imitation learning which learns to cede control to a supervisor when (1) in novel states or (2) in bottlenecks where task success is unlikely.

Safe and efficient RL from image observations by leveraging suboptimal demonstrations to structure exploration and examples of constraint violations to satisfy user-specified constraints.

A scalable and robust RL algorithm which optimizes for a combination of expected performance and tail risk under a distribution over learned reward functions.

An algorithm for safe reinforcement learning which utilizes a set of offline data to learn about constraints before policy learning and a pair of policies which seperate the often conflicting objectives of task directed exploration and constraint satisfaction to learn contact rich and visuomotor control tasks.

An MPC-based algorithm for robotic control (ABC-LMPC) with (1) performance and safety guarantees for stochastic nonlinear systems and (2) the ability to continuously explore the environment and expand the controller domain.

A new algorithm for safe and efficient reinforcement learning (SAVED) which leverages a small set of suboptimal demonstrations and prior task successes to structure exploration. SAVED also provides a mechanism for handling state-space constraints by leveraging probabilistic estimates of system dynamics.

A new formulation of imitiation learning from a non-stationary supervisor, associated theoretical analysis, and a practical algorithm to apply this formulation to develeop an RL algorithm which combines the sample efficiency of model-based RL and the fast policy evaluation enabled by model-free policies.