Kushal Kedia
kk837@cornell.edu
PhD Student
I am an incoming postdoctoral researcher at MIT, where I will be working with Prof. Pulkit Agrawal. I completed my Ph.D. in Computer Science at Cornell University, advised by Prof. Sanjiban Choudhury and Prof. Wei-Chiu Ma, during which I was also a visiting researcher working with Prof. Jeannette Bohg and Prof. C. Karen Liu at Stanford University.
My research goal is to build dexterous robots that can perform manipulation tasks which are difficult to demonstrate for humans. Towards this goal, I work on lifelong reinforcement learning algorithms that enable robots to acquire these capabilities from their own experience. I believe such robots will augment humans in everyday life and work.
Link to CV.
News
| Jun, 2026 | Our paper SimToolReal has been accepted to RSS 2026 and won a Best Poster award at the Dexterous Manipulation Workshop at ICRA 2026! SimToolReal learns an object-centric policy for zero-shot dexterous tool manipulation. |
|---|---|
| May, 2026 | I am organizing the Beyond Teleoperation workshop at ICRA 2026, imagining a future where robots are trained without any teleoperation data. |
| Feb, 2026 | Gave invited talks on SimToolReal at the Dexterity Team at NVIDIA, Google DeepMind, the ROAM Lab at Columbia University, and GRAIL at NYU. |
| Oct, 2025 | Gave invited talks on “Moving Beyond Teleoperation: Robot Learning from Human Videos” at the University of Michigan’s Computation HRI course (course link), the Robotics Mobility Team at NVIDIA, and the Foundation Models Team at the RAI Institute. |
| Aug, 2025 | Our paper X-Sim received an oral presentation at CoRL 2025 and a Best Paper (runner-up) award at the EgoAct Workshop at RSS 2025. X-Sim trains real‑world image‑based robot policies entirely in simulation, without any teleoperation data. |
| Apr, 2025 | Cornell Chronicle featured our ICRA 2025 paper on RHyME (Retrieval for Hybrid Imitation under Mismatched Execution): “Robot see, robot do: System learns after watching how-tos.” |
| Oct, 2024 | Gave invited talks on “Transferring Collaborative Behaviors from Human-Human Teams” at RPM Lab, University of Minnesota and RobIn Lab, UT Austin. |
| May, 2024 | Released MOSAIC, a year-long collaborative effort combining multiple foundation models to build a multi-robot collaborative cooking system. MOSAIC won the Best Paper award at the VLMNM workshop and the Best Poster at the MoMa workshop @ ICRA 2024! |
| Sep, 2023 | Excited to release the Collaborative Manipulation Dataset (CoMaD)! CoMaD captures over 6 hours of motion data from 14 unique users involving both human-human and human-robot (Franka robot arm) interactions in a kitchen setting. |
Selected Papers
-
Play2Perfect: What Matters in Dexterous Play Pretraining for Precise Assembly?arXiv 2026 -
SimToolReal: An Object-Centric Policy for Zero-Shot Dexterous Tool ManipulationRSS 2026
Best Poster @ Dexterous Manipulation Workshop, ICRA 2026 -
X-Diffusion: Training Diffusion Policies on Cross-Embodiment Human DemonstrationsICRA 2026 -
X-Sim: Cross-Embodiment Learning via Real-to-Sim-to-RealOral Presentation @ CoRL 2025
Best Paper (Runner-Up) @ EgoAct Workshop, RSS 2025 -
-
MOSAIC: Modular Foundation Models for Assistive and Interactive CookingCoRL 2024
Best Paper @ VLNMN Workshop, ICRA 2024
Best Poster @ MoMa Workshop, ICRA 2024 -
InteRACT: Transformer Models for Human Intent Prediction Conditioned on Robot ActionsICRA 2024 -
ManiCast: Collaborative Manipulation with Cost-Aware Human ForecastingCoRL 2023 -
