Position details

Purpose of the role

Lead the Edge AI Optimization Department in taking Physical AI models (VLA, world models, RL policies) from research to real robots: running in real time, reliably and energy-efficiently on low-cost edge hardware. The role owns the full optimization and deployment toolchain (VLAOpt), from model compression to inference runtime and integration on VinRobotics humanoid platforms.

Key responsibilities

1.   Technical strategy: build and execute the edge optimization and deployment roadmap for the robot platforms; work with the Reinforcement Learning and Robot Foundation AI departments so new models are deployable by design.

2.   Runtime and toolchain: lead development of the inference runtimes (vla.cpp, vla.simd) and hardware backends (CUDA/TensorRT, Qualcomm Hexagon NPU, Intel OpenVINO, CPU SIMD); productize the VLAOpt toolchain.

3.   Model compression: low-bit quantization (FoldQuantVLA), pruning, distillation and sub-1B-parameter VLAs; balance latency, accuracy and BOM cost.

4.   Field deployment: meet latency and reliability targets for factory use cases; work with the EE team to benchmark SoCs/SOMs and recommend hardware choices.

5.   Partners and ecosystem: technical point of contact with NVIDIA, Qualcomm and Intel on model support, dev kits and porting; maintain open-source releases (GitHub, Hugging Face), publications and patents.

6.   Team leadership: hire to headcount, run the KPI process, and mentor engineers and talent-program members (interns, Young Engineers, Residents).

Requirements

•     MSc or PhD in Computer Science, Computer Engineering, Electronics or a related field.

•     5+ years optimizing and deploying deep learning models on edge devices or in production, including 2+ years leading an engineering team.

•     Strong modern C++ and Python; deep understanding of hardware architecture (GPU, NPU, SIMD NEON/AVX, memory hierarchy).

•     Hands-on experience with at least two of: TensorRT, ONNX Runtime, OpenVINO, Qualcomm QNN, TVM, ggml/llama.cpp, ExecuTorch.

•     Solid command of quantization (PTQ, QAT, INT8/INT4), pruning, distillation and transformer inference optimization (KV cache, kernel fusion, batching).

•     Professional working English for direct collaboration with international partners.

Preferred

•     Has deployed VLAs, VLMs or LLMs on robots or embedded devices; understands diffusion/flow-matching action heads.

•     CUDA or NPU kernel development; experience with real-time systems and ROS 2.

•     Publications at top venues (CoRL, ICRA, RSS, NeurIPS, MLSys) or impactful open-source contributions.

•     Experience evaluating hardware and optimizing BOM cost for commercial products.