Research Engineer (Robotics) · Robot Learning & Embodied AI

Chandramouli Rajagopalan

I build data-efficient robot learning systems that transfer from simulation and offline data to real-world robots, then continue adapting under changes in dynamics, perception, and task conditions.

Sim-to-Real Domain Randomization Adaptive Learning Continual Learning Offline RL World Models

Selected Research Publications CV

GitHub LinkedIn Email

My research focuses on robust robot learning under distribution shift. I study how simulation, offline data, uncertainty-aware adaptation, and multimodal feedback can be combined to initialize policies efficiently and improve them from limited real-world experience. My long-term goal is to develop robots that adapt to new dynamics, observations, tasks, and embodiments without forgetting previously learned skills.

Robust Sim-to-Real Transfer

Develop policies that tolerate inaccuracies in dynamics, contact, latency, perception, and hardware interfaces.

Adaptive and Continual Robot Learning

Enable robots to detect distribution shift, adapt from limited real-world interaction, and acquire new skills without catastrophically forgetting prior capabilities.

Data-Efficient Real-World Learning

Use offline RL, world models, imitation learning, and pretrained multimodal models to reduce expensive real-robot interaction.

Robust Sim-to-Real Manipulation

How can policies trained in simulation retain precision under contact variation, hardware mismatch, latency, and imperfect observations?

Built end-to-end sim-to-real reinforcement-learning pipelines for object stacking and peg insertion on OpenArm, as well as free-grasp transfer on the Panthera-HT manipulation platform. The work covered simulation task design, contact-rich policy training, observation and action interface alignment, hardware integration, and real-robot evaluation.

My contribution: Simulation task design, contact-rich policy training, observation/action interface alignment, hardware integration, and real-robot evaluation.

Occlusion-Robust Ping-Pong Control

30 minutes uninterrupted

How can a dynamic control policy remain stable when state observations are temporarily unavailable or noisy?

Developed and deployed a state-based sim-to-real controller for dynamic robotic ping-pong ball control. The system sustained continuous paddle-ball juggling on real hardware for 30 minutes without failure while maintaining and recovering closed-loop state estimates through temporary ball occlusions, observation noise, and timing-sensitive contacts.

30 minutes uninterrupted
Real-robot deployment
Recovery through temporary occlusion

My contribution: State-based sim-to-real control, real-hardware deployment, and robustness evaluation under occlusion, observation noise, and timing-sensitive contacts.

Offline RL and World-Model Adaptation

How can offline robot data provide a strong initialization while limited online interaction corrects distribution shift safely?

Contributed to Finetuning Offline World Models in the Real World, which pretrains world models using offline real-robot data and subsequently adapts them using limited online interaction and uncertainty-aware planning.

Developed Value-Guided Action Transport, an offline reinforcement-learning method that refines behavior-cloned actions using Q-guided residual transport, path-based value evaluation, adaptive advantage gating, and a behavior-cloning anchor. The method achieved the highest or tied-highest normalized score on 9 of 12 evaluated D4RL tasks in the reported experiments.

Value-Guided Action Transport
Ongoing Research

My contribution: Offline-to-online world-model adaptation research contributions and development of Value-Guided Action Transport for offline reinforcement learning.

Paper Project

Vision-Language Robot Learning

Can pretrained multimodal models provide scalable supervision for long-horizon robot learning?

Designed reinforcement-learning environments for XLeRobot and BrainBot and fine-tuned Vision-Language-Action policies for robotic manipulation tasks. Developed a VLM-based progress-scoring system that estimates partial task completion from visual observations, providing scalable feedback for policy evaluation and long-horizon robot-learning experiments.

My contribution: Reinforcement-learning environment design, VLA policy finetuning, and VLM-based progress scoring for long-horizon manipulation experiments.

Adaptive Domain Randomization

Can a robot use real-world failures to identify which simulator parameters are incorrect and automatically update its randomization distribution?

Safe Offline-to-Online Adaptation

Can uncertainty, value estimates, and distribution-shift detection determine when a deployed policy should adapt and how much real-world exploration is safe?

Continual Learning Across Tasks and Embodiments

How can a robot acquire new manipulation skills, environments, and hardware embodiments while preserving previously learned behaviors?

Finetuning Offline World Models in the Real World

Yunhai Feng, Nicklas Hansen, Ziyan Xiong, Chandramouli Rajagopalan, Xiaolong Wang

Conference on Robot Learning (CoRL), 2023 — Oral

Offline real-robot world-model pretraining followed by uncertainty-aware online finetuning for data-efficient adaptation to seen and unseen tasks.

Paper Project

Deep Learning in a Bilateral Brain With Hemispheric Specialization

Chandramouli Rajagopalan, David Rawlinson, Elkhonon Goldberg, Gideon Kowadlo

Springer LNCS / ACAIN 2024

A bilateral neural architecture in which complementary local and global feature specializations are integrated through a learned network head.

Paper Code

Universal Domain Adaptation to Improve Performance on Out-of-Distribution Data

PMLR / NeurIPS 2021

Additional experience, projects, and technical details are available in my CV.

I am a robotics research engineer working on sim-to-real transfer, reinforcement learning, world models, and multimodal robot learning. I received an M.S. in Intelligent Systems, Robotics and Control from the University of California, San Diego.