Research and develop techniques to enable real-time inference for embodied AI models deployed on robotic platforms.

Optimize inference performance for models such as:

Vision-Language-Action (VLA) models

World models

Multimodal transformer-based policies

Perception and state estimation models used in robot control loops

Improve model latency, throughput, memory efficiency, and system reliability through methods such as:

model compression

quantization

distillation

batching and scheduling optimization

KV-cache / decoding optimization

graph compilation and kernel-level acceleration

Collaborate with robotics, infrastructure, and hardware teams to integrate optimized models into real robot stacks and edge/on-device systems.

Design benchmarking pipelines for evaluating end-to-end performance, including control frequency, action latency, and system robustness under real deployment constraints.

Explore tradeoffs between model quality and runtime efficiency to support practical deployment in real-world robotic tasks.

Contribute to internal technical reports, system design discussions, and publications where appropriate.

Currently pursuing or recently completed a PhD in Computer Science, Electrical Engineering, Robotics, Machine Learning, Systems, or a related field.

Strong background in machine learning systems, model inference optimization, or efficient deep learning.

Experience optimizing modern ML models for production or low-latency deployment.

Hands-on experience with one or more of the following:

real-time inference systems

efficient transformer inference

model compression, pruning, quantization, or distillation

GPU performance optimization

deployment frameworks such as TensorRT, ONNX Runtime, XLA, TVM, Triton, or similar systems

Proficiency with deep learning frameworks such as PyTorch, JAX, or TensorFlow.

Strong programming and systems skills, including experience with performance profiling and debugging.

Ability to work across the stack, from model architecture to runtime systems and hardware-aware optimization.

Requires 5 days/week in-office collaboration with the team.