Ajinkya Jain

Hi! I am a Senior Research Scientist at Google (Intrinsic), where I am part of the R&D team led by Prof. Stefan Schaal. My research lies at the intersection of robot learning, embodied foundation models, and physics-informed control, focusing on methods that enable robots to perform dexterous, contact-rich, and multi-robot manipulation reliably in the real world. Recently, my work has centered on on-policy distillation and reinforcement learning finetuning, policy synthesis with foundation models (VLMs/LLMs), and force-aware imitation and online learning for tight-tolerance assembly.

Previously, I was a Robotics Engineer on the R&D team at Vicarious (acquired by Alphabet in 2022 and integrated into Intrinsic). I received my Ph.D. in Robotics from The University of Texas at Austin under the guidance of Prof. Scott Niekum and Prof. Ashish Deshpande, where my doctoral research focused on learning object articulation and hybrid interaction models from visual data and motion planning under uncertainty. Prior to UT Austin, I earned my Dual Degree (B.Tech.–M.Tech.) in Mechanical Engineering (Robotics and Automation) from IIT Kanpur.

Email  /  GitHub  /  Google Scholar  /  LinkedIn  /  Resume  /  CV

Research

My research develops scalable learning and control algorithms for generalist, contact-rich robotic manipulation—spanning reinforcement learning finetuning and policy distillation, foundation models (VLMs/LLMs) for policy synthesis, force-aware imitation and online learning, sim-to-real transfer, and motion planning under uncertainty. Alongside my publications and patented industrial manipulation technology (US Patent 12,330,304 B2), I co-organize community workshops at RSS 2026, IROS 2026, and IEEE Humanoids 2026 on sim-to-real transfer and learning-based whole-body robot control.

RoboBallet2: Multi-robot Reaching for Arbitrary Geometries via On-policy Distillation and Reinforcement Finetuning

Scaling collision-free multi-robot reaching and motion generation across arbitrary 3D obstacle geometries via on-policy distillation from graph-neural-network experts and reinforcement learning finetuning.

Predictive Force Conditioning for Contact-Rich Imitation Learning

A two-stage imitation learning architecture that explicitly forecasts future force-torque dynamics and conditions action generation on anticipated physical contact feedback for fine-grained, tight-tolerance manipulation.

Efficient Online Learning of Contact Force Models for Connector Insertion

An efficient and scalable approach for modeling connector insertion environments by learning a quasi-static contact force model instead of a full simulator, employing a Linear Model Learning algorithm for real-time optimal mapping without matrix inversions.

RT-Sketch: Goal-Conditioned Imitation Learning from Hand-Drawn Sketches

A goal-conditioned manipulation policy that uses hand-drawn sketches for intuitive, spatially grounded, and distractor-robust goal specification in tabletop and articulated rearrangement tasks.

GenCHiP: Generating Robot Policy Code for High-Precision and Contact-Rich Manipulation Tasks

Equips large language models with compliant impedance and force-constrained action primitives to synthesize zero-shot robot policy code for high-precision, tight-tolerance contact-rich manipulation tasks.

A Comparison of Imitation Learning Algorithms for Bimanual Manipulation

A systematic empirical evaluation of modern imitation learning algorithms (Diffusion Policy, ACT, IBC, GAIL, DAgger, and Behavior Cloning) on over-constrained, high-precision bimanual insertion tasks.

Open X-Embodiment: Robotic Learning Datasets and RT-X Models

Large-scale datasets in standardized data formats and models to enable learning “generalist” X-robot policy that can be adapted efficiently to new robots, tasks, and environments.

Distributional Depth-Based Estimation of Object Articulation Models

A novel method that efficiently learns distributions over articulation model parameters directly from depth images without the need to know articulation model categories a priori

ScrewNet: Category-Independent Articulation Model Estimation From Depth Images Using Screw Theory

A novel method for estimating articulation models for objects directly from raw depth images without knowing their articulation type a priori using screw theory

Learning Hybrid Object Kinematics for Efficient Hierarchical Planning Under Uncertainty

A method for learning planning-compatible hybrid kinematics models for articulated objects from human demonstrations.

Efficient Hierarchical Robot Motion Planning Under Uncertainty and Hybrid Dynamics

A POMDP-based Motion planner that generates efficient motion plans leveraging object-object interactions to perform manipulation tasks under uncertainty with high success rates

Belief Space Planning under Approximate Hybrid Dynamics

Extending Belief-space LQR to hybrid systems for robot motion planning under uncertainty

A piezoelectric model based multi-objective optimization of robot gripper design

Multi-objective design optimization of a piezoelectric actuator driven gripper

Unified Minimalistic Modelling of Piezoelectric Stack Actuators for Engineering Applications

A simple, computationally-cheap, yet effective model for piezoelectric stack actuators as a replacement of black-box models used in engineering design optimization problems.

Design and source code from Jon Barron's website. Jekyll version from Leonid Keselman's website