Vakhtang “Vato” Inasaridze

Robot learning, with the hardware to make it real.

Mechanical engineer, First Class from Warwick, admitted to UC Berkeley MEng (Robotics). I build robot learning and 3D perception systems, and the mechanisms, PCBs and firmware to run them.

Selected work

HardwareBEng dissertation · solo · 2025-26

MRV: Monopedal Reconnaissance Vehicle

A one-legged jumping robot for low-gravity exploration, stabilised by two inertial actuators. Self-proposed; I designed every mechanical part, the 4-layer PCB and the Teensy C++ firmware (120 Hz sensor fusion, independent 1 kHz safety interrupt).

  • Five revisions of the spring release mechanism; asymmetric ratchet wedge stored 38% more energy per cycle.
  • Root-caused a 43% launch failure rate to a cam concentricity offset.
  • CAD, firmware and BOM open under MIT licence.
41 cmmean jump at 360 g
67.8%elastic to potential energy
12.7°attitude recovery
£286of £300 budget
github.com/termvato/MRV →
Robot learningOct 2026

Phone-to-robot imitation learning

iPhone LiDAR recordings of my hand turned into robot demonstrations in LIBERO: hand tracking, metric 3D lifting, grasp-anchored retargeting and a demo multiplier.

41%ACT from 3 phone takes (92% scripted)
3/4pre-registered world-model tests passed
Code →
3D perceptionOct 2026

3D room change detection

Two phone videos in; moved, added and removed objects out, named, with ROS occupancy maps. Hosted GPU web app on Modal.

2.4→18.0AP with free-space voxel carving
21.4typed AP vs 15.8 SceneDiff (ECCV 2026)
Code →
ControlOct 2026

Reaction-wheel pendulum: LQR vs PPO

Isaac Sim plant, LQR as a ROS 2 node, PPO trained in Isaac Lab as a drop-in replacement. A linear program over all torque sequences proved no controller beats 7.47°; both reached 7.45°.

Code →
Medical deviceTeam of 5

Cannulation trainer

Forearm simulator for dialysis nurses on a £200 budget. I owned the flesh analogue: silicone matched to forearm Shore hardness, vein found by touch, replaceable vein. Chosen best of five teams by the nurses.

Awards

1st place, Physical AI Hack London

60 builders. Rocky Robotics: quarry autonomy. Simulation, world generation, voice fleet commands, business case.

Top 8 of ~180, Accel AI Innovate London

CounterFactual: crash recovery decisions for insurers. Simulator, careful-driver models, CCTV vision, agent team.

European Physics Olympiad, represented France

Hannover. BPhO Intermediate Challenge Gold.

Experience & education

UC Berkeley, MEng Mechanical Engineering (Robotics)

Admitted, deferred.

University of Warwick, BEng Mechanical Engineering

First Class Honours. Robotics Society Project Manager: membership 15 to 28.

Siemens, Manufacturing Automation Intern

Specified a £400k selective soldering machine and its capital case; approved by six chief engineers and installed.