Foundation Model Robotics — Research & Development
I research foundation models for embodied agents — studying how vision-language-action policies perceive, generalize, and act reliably in the physical world.
About
My work sits at the intersection of foundation models and embodied intelligence. I study how vision-language-action (VLA) models can be trained, evaluated, and made reliable enough to control real robots — with a particular focus on understanding where and why these policies break down before trusting them in the physical world.
I work across the full pipeline: generating training data in simulation with NVIDIA Isaac Lab, benchmarking VLA policies under realistic distribution shift, and making large vision-language models efficient enough to deploy through quantization and inference optimization. I care as much about rigorous, reproducible evaluation as I do about headline results.
Experience
Working directly alongside the core engineering team to develop multi-modal vision-language-action (VLA) models, design and test robotic interfaces, and contribute to physical hardware prototyping — hands-on, AI-driven automation within a real-world wet-lab development cycle.
Research
Three threads of my foundation-model robotics work — from generating data in simulation, to stress-testing policies, to making large models efficient to run.
A full sim-to-action pipeline in NVIDIA Isaac Lab: deterministic Warp state machines and a PPO teacher (4096 parallel envs) generate Franka pick, barrier, and push demonstrations; multimodal streams (overhead RGB-D, wrist RGB, joint kinematics, contact forces) are recorded and converted from HDF5 to LeRobot v3.0, then used to train ACT policies under rigorous closed-loop evaluation with camera ablations and out-of-distribution tests.
Barrier 90% (>75% oracle) · diagnosed a 0% push policy's camera shortcut
How well do VLA policies hold up when the robot starts somewhere new? I benchmark a monolithic SmolVLA policy against a hierarchical Gemini + IK system on tabletop pick-and-place in MuJoCo, across controlled spatial shifts in the initial pose.
Degrades beyond ~10 cm offset · 4 shift conditions
A failure-mode study of LLM quantization applied to a vision-language model. Standard FP8 saturates the visual-token KV cache and collapses accuracy; SmoothQuant (W8A8) is Pareto-optimal — roughly 2× faster while retaining accuracy.
≈2× speedup · 7 configs evaluated