Hardware & Physical Systems
Before working on learned models, I spent several years doing hardware-first systems work. This background shapes how I think about reliability, failure modes, and the gap between simulation and deployment — problems that remain unresolved in modern physical AI.
Nino: Consumer Humanoid Robot — Sirena Technologies (2014–2016)
Founding engineer on Nino, a consumer humanoid robot developed at Sirena Technologies in Bangalore. Led the complete mechanical and control architecture from concept to walking prototype.
Mechanical design. Defined the full kinematic structure: degree-of-freedom allocation, joint placement, and actuator sizing via static load analysis to optimize stability margins. The tradeoffs here — between stability, weight distribution, actuator torque budget, and gait flexibility — are not separable from the control problem. Getting DoF selection wrong means no control algorithm can compensate.
State estimation. EKF-based sensor fusion across IMU, force/torque sensors, and encoders. The practical challenge: sensor noise characteristics change with gait phase, and the filter needs to remain reliable across contact events that introduce impulsive dynamics.
Gait control. 3D Linear Inverted Pendulum (3D-LIP) model for gait generation and real-time balance control. Hierarchical control stack connecting high-level gait planning to low-level joint impedance control.
Systems integration. Owned the hardware-software interface across embedded motor drivers, sensor buses, and high-level planning. The lesson from this work: reliability in physical systems is a design property, not something you can retrofit. It requires reasoning about failure modes before they happen.
Neuro-Adaptive Optimal Control — Quadcopters (IISc)
Research on adaptive control for quadrotors at the Indian Institute of Science (IISc), where the goal was maintaining stable flight under model uncertainty and external disturbances. The setup: a nominal optimal controller paired with a neural network that continuously learns unmodeled dynamics online and updates the control policy in response: simultaneous online model identification and policy adaptation.
In RL terms: the agent is updating both its world model estimate and its policy in closed loop, with no episode resets and finite horizon replanning. The interesting tension is that the neural component improves performance under disturbances but trades away the clean stability guarantees of purely classical methods — a concrete instance of the expressivity-vs-guarantees tradeoff that appears throughout learned control.
Satellite Systems — ISRO Collaboration
First introduction to systems engineering and safety-critical hardware design as part of a satellite team collaborating with ISRO. The focus was fault-tree analysis: systematically enumerating failure modes, estimating their probabilities, and designing redundancy structures to meet reliability targets.
What this taught me that ML research often misses. In safety-critical systems, reliability is specified before you build the system, not measured after deployment. The design process starts with a failure budget and works backwards to component-level requirements. This is almost the opposite of how most ML systems are deployed: trained, shipped, and then evaluated on failures as they occur.
The satellite work established an intuition I’ve returned to repeatedly: the relevant question is not just “does this system work?” but “under what conditions does it fail, and how detectably?” That question is now central to my research on uncertainty and representation.