Hi, I'm Adithya
I also get them running on the robot itself: small chips, tight power budgets, real-time deadlines.
Robots used to be programmed. Now they're starting to be taught.
Show a robot enough demonstrations, give it a camera and a sentence, and it can learn to act. I'm curious about what comes after that — when robots don't just see and listen, but also feel, anticipate, and work safely next to people.
Vision-Language-Action models that turn pixels and words into motion.
Giving policies a sense of touch, for the delicate, contact-rich tasks that vision alone can't handle.
Fitting big models onto small computers, so a robot can decide what to do in under 30 milliseconds.
Certifiable, reproducible AI for machines that share space with humans.
How I got here
I studied automotive engineering and found that the part I loved most was the software deciding what the machine should do. At Accenture and EMIS Health I worked on telecom data, speech models for NHS emergency calls, and the data lake that fed them.
a model is only as good as the pipeline behind itA master's in Advanced Robotics at École Centrale de Nantes, where I built my robotics foundations: kinematics, dynamics, control, sensors, AI and humanoids. My favourite project used motion capture and force plates to have a humanoid copy a person's upper body in real time.
first time a robot moved because I movedAt Reliev I worked with ECG, EEG and other bio-signals to detect epileptic seizures, with deep learning on top of careful signal processing. I also built the pipeline that handled ECG data at scale, from ingestion and cleaning to training, all aimed at making life safer for people living with epilepsy.
same question as robotics: what is this signal trying to tell me?At Huber Automotive I built camera, LiDAR and radar perception for autonomous vehicles on tight embedded budgets, and worked with Fraunhofer IPA on VeoPipe, a pipeline for AI you can actually certify.
At Neura Robotics I fine-tune Vision-Language-Action models on teleoperation data, give them a sense of touch, and get them running on Jetson and Qualcomm hardware fast enough to close a real control loop.
this is the fun part · see Neuron AI ↗Here are some things I've been working on lately…
A profiling toolkit for running NVIDIA's GR00T N1.5 on Jetson. It breaks inference latency into the vision encoder, the language backbone and the diffusion action head, so you can see where the milliseconds go.
What work gloves do to the 3D hand tracking behind human-video robot learning, and which capture choices fix it: a second camera for hand position, a thin glove for finger pose.
Working with me
I'm happiest on problems where a model works in a demo but not yet in the field. I usually deliver two things: the working system, and the measurements that explain what was going wrong before. When I'm exploring a new team, these are the questions I like to ask.
I lean towards shipping early and learning from real use, but I care about craft too. I'd like to know where the balance sits.
Whether failures get traced back to a root cause and shared openly, or quietly patched and forgotten. How a team handles failure says a lot about it.
I do my best work when I'm responsible for an outcome, not just a list of tasks.
Whether senior engineers make time for others, and whether there's room for me to mentor someone who could one day take over my work.
Some of the best fixes come from stepping back and looking at the whole pipeline. I'm curious which habits, reviews and rituals are deliberate here.
I'm always happy to chat, whether that's about robots, research, or interesting problems.
Say hello →I speak Tamil at home, English at work, French when I'm feeling brave, and German… ich übe noch.
