Elijah Tabachnik — Machine Learning Engineer · RL & Simulation Systems · Irvine, CA
Reinforcement learning, robot simulation, and the unglamorous systems that make both ship. UC Irvine CS, June 2026.
Continual reinforcement learning under hidden distribution shifts — measured against the control ladder most work skips.
An outer PPO “Brain” meta-controller sets an inner continual learner’s plasticity online — learning rate, replay, anchoring, and a neuromodulatory context code — while the environment’s reward regime flips with no observable cue. Run against matched controls, the trained Brain yields a small but statistically real recovery gain, while a zero-forgetting oracle reveals eight times more headroom. The bottleneck isn’t learning dynamics — it’s knowledge restoration.
hit-rate-80 — fraction of hidden shifts recovered to ≥80% success
trained-vs-init recovery gain (composite, p<0.01, n=32 eval seeds)
headroom revealed by a zero-forgetting oracle — memory is the bottleneck
RL locomotion and LiDAR-guided navigation trained across 4,096 parallel Unitree Go2 simulations, alongside the mechanical design to eventually build it.
parallel Go2 environments during training
The design track runs alongside the autonomy stack: actuator selection and torque margins, fused power distribution, and a single-leg validation loop in Isaac Sim + ROS2 before any hardware gets risked. Reward shaping mattered as much as architecture — early policies found every poor-but-legal shortcut the simulator allowed.
| P-03 |
SkyPilot — Autonomous Drone Voice-commanded autonomy: Whisper → LLM intent routing → YOLOv8/OpenCV tracking. |
CV · LLM | → |
| P-04 |
Procedural Terrain Generator Layered Perlin noise, biome foliage, 15K+ objects with LOD — Unity/C#. |
Graphics | → |
| P-05 |
Reviving History Historical-persona simulation: LLM conversation, voice synthesis, avatar interaction. |
Multimodal | → |
I’m a machine learning engineer focused on reinforcement learning and simulation systems — currently finishing a CS degree at UC Irvine (June 2026) and running a controlled research program on continual learning. I like problems where the environment fights back: regime shifts, sim-to-real gaps, and reward functions that get exploited in ways nobody predicted.
Education: B.S. Computer Science, University of California, Irvine — expected June 2026. Coursework in machine learning, artificial intelligence, algorithms & data structures, and operating systems.
Open to machine learning, robotics, and applied AI roles starting 2026 — say hello.