Postdoctoral Researcher, Dynamic Legged Systems lab,
Istituto Italiano di Tecnologia (IIT), Genova, Italy
I build legged and modular robots for the lunar surface, and the learning-based control that lets them keep working when the terrain does not cooperate.
I am a postdoctoral researcher in the Dynamic Legged Systems lab at Istituto Italiano di Tecnologia (IIT) in Genova, working with Claudio Semini on learning-based locomotion controllers for quadrupedal robots in space exploration, under the ALTEC project. Most of my time goes into GPU-accelerated reinforcement learning and the sim-to-real transfer pipelines that decide whether a policy trained in Isaac Lab survives contact with real hardware.
Before Genova I completed my doctorate at the Space Robotics Lab at Tohoku University with Kazuya Yoshida, as a Japanese Government MEXT scholar, working on MoonBot: a modular, reconfigurable robot for lunar base construction developed under JST Moonshot Goal 3. That work was field-tested at the JAXA Space Exploration Test Field in Sagamihara and in the Abukuma Cave, which we used as a stand-in for lunar cave environments.
IEEE Transactions on Field Robotics, 2025
One set of limbs, wheels and tools that reconfigures on demand into whichever machine a lunar construction task needs.
Best Paper Award in Planetary Robotics, IEEE iSpaRo 2025, Sendai, Japan
IEEE International Conference on Space Robotics (iSpaRo), 2025
A decentralized reinforcement learning approach for robots whose morphology changes: each module learns to act on local, multi-modal observation rather than from a central controller, so the same policy set survives reconfiguration.
My research lies at the intersection of space robotics and artificial intelligence.
Current work
How does a legged robot keep walking on terrain that no one has surveyed, with no chance of rescue?
Learning-based controllers for quadrupeds that have to walk on surfaces nobody has surveyed.
If you can only land a fixed mass on the Moon, should it be one specialised machine or a set of parts?
MoonBot: one set of limbs, wheels and tools that reconfigures on demand into whichever machine a lunar construction task needs.
What has to be true for a learned policy to still work outside the simulator?
Decentralized RL, agentic planning, and the sim-to-real gap that decides whether any of it survives contact.
IAC 2025 in Sydney, a best paper award at iSpaRo 2025, and the Space Studies Program in Houston.
The JST Moonshot Goal 3 programme behind MoonBot: robot systems that can build and maintain a lunar outpost with minimal human help.
JAXA Space Exploration Test Field, Sagamihara
Full-scale field validation of agentic planning and execution for modular lunar robots, in lunar-analogue terrain and lighting.
JAXA Space Exploration Test Field, Sagamihara
Reinforcement learning control policies tested on morphology-varying modular robots across multiple reconfigurations.
Abukuma Cave, Japan
A modular sensor package for the MoonBot platform, fusing LiDAR, fisheye and depth cameras. Field tested by mapping a cave.
10 peer-reviewed items in total, plus work under review.
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
Keeping multiple limbs of a space robot synchronized when any of them may be added, removed or delayed, using hypersphere trajectory clamping to bound the motion each limb is allowed to command.
doi:10.1109/IROS60139.2025.11246735arXiv preprintVideoCode: Motion Stack
IEEE International Conference on Automation Science and Engineering (CASE)
Human-in-the-loop verification and validation of UR16e manipulator control for precise assembly and deployment under uncertain terrain and visual conditions, using ArUco markers and OpenCV.
Outside the lab I am Technical Lead and Advisor at Helping Hut, a student-founded NGO running free street schooling, a village library and relief drives in India, and I built the platform it runs on.
Collaboration enquiries, talk and panel invitations, and questions about the work are all welcome. I read every message. Write to contact@drmishra.space or use the contact page.