Projects by Nikhil Prabhu — robotics, reinforcement learning, and agent systems

I made 19 frontier models run a restaurant, built a playable version to collect human traces for post-training, live-demoed it at True Ventures' Scrappy Sessions, and separately presented it to Mercor's Code Data team at Mercor HQ.

♦1st Overall @ Bow Capital × DS3 × SIC Hackathon

End-to-end autonomous drone platform that identifies targets, navigates without GPS, and coordinates swarm behavior without communication, we won 1st Overall at the Bow Capital × DS3 × SIC hackathon @ UC San Diego.

♦1st Overall @ DiamondHacks 2026

Voice-first AI companion for senior care: an autonomous robot with natural speech interaction, wake-word detection, medication reminders, crisis detection, and a caregiver monitoring dashboard. Frontend built with Svelte 5 and Bun runtime for a lightweight, responsive experience.

♦1st Overall @ OpenClaw Hackathon Eragon x Nozomio x AgentMail

RL training orchestrator for a 2D Roomba cleaning sim that automates the full workflow: FastAPI trains PPO agents, evaluates them, generates GIFs, and delivers results through email (AgentMail), memory (Nia), and MCP tools for autonomous agent-driven optimization.

♦1st Overall @ transpose/compiled-5 UCSD Hackathon

Navigation, Incident Guidance, and Emergency Localization & Control — a mission-control interface for the Future Interfaces hackathon. Think air traffic control for firefighters: helmet VSLAM point clouds, multi-unit video feeds, radio transcripts, and tactical mapping in one dispatcher dashboard.

♦5th Place @ Nozomio Hackathon

Deploy GPUs with natural language and find cost arbitrage across cloud providers. A control plane that routes inference workloads across local Docker, Kubernetes, Vultr, Vast, and llama.cpp while automatically finding the cheapest, fastest hardware matches for your models.

Rescue mission control for simulated disaster response. PPO training on eight MuJoCo simulation environments, Gemini environment generation, and a language-driven console with live rollouts and GIF replay.

A pre-execution world model verifier for robot skills. Claude proposes a pick-and-place plan, a learned action-conditioned world model imagines the outcome from camera frames, and a verifier predicts failure so the plan gets repaired before the arm ever moves.

Teaching a Unitree G1 humanoid to score a penalty kick in MuJoCo. Human penalty-kick motion is retargeted through a pelvis-local frame, then a differentiable proxy warm-starts a CEM search that turns a 3.7 m dribble into accurate shots to both corners.

Object-centric imitation policy that augments ACT with YOLOv8 detections for pick-and-place under visual corruption.