Skip to main content
TALENT247 // SPECIALIST AI EXECUTIVE SEARCH & TECHNICAL RECRUITMENT
ROB · Robotics & Embodied AI

Senior Real-Time Robotics & On-Device Inference Engineer

San Francisco, CA · Full-time · On-site

AI Specialism

Robotics & Embodied AI

Workplace

On-site

Location

San Francisco, CA

Compensation

$225k – $320k

annual package

Role Dossier & Specification

REF: SENIOR-REAL-TIME-R

THE CLIENT & MANDATE

Our client is a premier San Francisco generalist-robotics AI lab training multi-billion-parameter Vision-Language-Action (VLA) foundation models to perform complex dexterous manipulation in unstructured physical environments. Turning massive multimodal transformers into smooth, reactive 50Hz–100Hz bimanual robot motion requires tightly co-designed C++ real-time control loops and ultra-low-latency GPU inference.

They have partnered with Talent247 to hire a Senior Real-Time Robotics & On-Device Inference Engineer onsite in San Francisco. Collaborating directly with VLA model architects and hardware engineers, the appointed engineer will own the onboard execution stack bridging raw multi-camera streams, TensorRT/CUDA action-chunk inference, and low-level torque/impedance controllers.

KEY RESPONSIBILITIES

  • Architect deterministic C++20 and CUDA/TensorRT inference engines running multi-billion-parameter VLA backbone + flow-matching action heads within <15ms end-to-end sensor-to-actuator latency.
  • Design asynchronous dual-rate execution architectures (decoupling heavy VLM KV-cache updates from high-frequency flow-matching action-chunk denoising and real-time trajectory interpolation).
  • Implement smooth temporal ensembling, receding-horizon spline blending, and compliant joint/Cartesian impedance controllers over CAN-FD, EtherCAT, and real-time Linux (PREEMPT_RT) buses.
  • Optimize custom CUDA/Triton kernels, CUDA Graphs, FP8/BF16 quantization, and zero-copy V4L2/CSI camera DMA pipelines on workstation and embedded edge GPUs.
  • Build high-precision hardware telemetry, safety envelopes (collision/torque watchdogs), and sub-millisecond synchronized episode recorders across dozens of daily robot stations.
  • Diagnose subtle sim-to-real and train-to-deploy discrepancies caused by camera exposure latency, clock skew, or numerical precision drift.

TECHNICAL ENVIRONMENT & OFFERING

Stack: Modern C++20, CUDA, TensorRT, Python, JAX/PyTorch export (ONNX/TorchExport/StableHLO), Linux PREEMPT_RT, EtherCAT/CAN-FD, multi-arm & mobile manipulator hardware. Workplace: Onsite in San Francisco, CA. Compensation: $225,000 – $320,000 base salary plus meaningful equity ownership and full medical/401(k) benefits.

Core Requirements

ESSENTIAL QUALIFICATIONS

  • 5+ years of robotics systems, real-time control, or low-latency GPU inference engineering experience in modern C++ (C++17/20) and Python.
  • Proven experience optimizing deep neural network inference on NVIDIA GPUs using CUDA, TensorRT, CUDA Graphs, and mixed-precision quantization.
  • Strong command of real-time Linux systems (PREEMPT_RT), lock-free concurrency, hardware sensor synchronization, and robotic kinematics/dynamics.
  • Hands-on track record bringing up and debugging real physical robot arms, dexterous hands, or mobile manipulators.
  • Obsessive attention to p99.9 latency jitter, safety interlocks, and clean hardware abstraction interfaces.

PREFERRED / BONUS EXPERIENCE

  • Direct experience deploying diffusion policies, flow-matching action heads, or VLA transformers on bimanual manipulation platforms.
  • Familiarity with exporting JAX/XLA or PyTorch 2.x custom attention graphs cleanly to C++/TensorRT runtimes.
  • Experience with whole-body operational space control, force/tactile feedback integration, or custom high-bandwidth motor drivers.

Position Telemetry

Location San Francisco, CA
Department Robotics & Embodied AI
Employment type Full-time
Workplace type On-site
Salary range
$225k – $320k annual package