Chưa ghi lương
Senior
5+ YoE
Hanoi
Xây dựng kiến trúc thị giác và triển khai mô hình Vision-Language lên robot nhân hình VinDynamics chạy thời gian thực trên edge.
Tóm tắt bằng AI từ JD gốc, đọc JD gốc trước khi ứng tuyển.
At VinDynamics, we design safe, affordable, and intelligent humanoid robots to assist in everyday life — robots for everyone. Backed by Vingroup, Vietnam’s leading technology conglomerate, we are on a mission to make advanced robotics accessible, reliable, and beneficial for billions of people worldwide. By combining cutting-edge AI, world-class engineering, and human-centered design, we aim to seamlessly integrate robots into daily life — enhancing safety, productivity, and happiness at home and beyond.
As our Computer Vision Expert, you will give our humanoid robots the ability to see, understand, and act on the physical world — turning frontier vision-language research into perception that runs in real time on the robot.
Vision-Language Perception & Open-World Understanding
• Design and own the visual perception architecture for humanoid robots, from raw multi-camera input to semantic scene understanding.
• Build and adapt Vision-Language Models (VLMs) for robot-centric tasks: visual question answering, scene description, referring-expression grounding, and instruction-conditioned perception.
• Develop open-vocabulary “detect anything / segment anything” pipelines that detect, segment, and track arbitrary objects in unstructured indoor environments without per-class retraining.
• Fine-tune, adapt, and distill vision foundation models on in-house robot data to improve robustness against lighting changes, occlusion, clutter, and motion blur.
Depth Sensing & 3D Perception
• Process and fuse RGB-D and stereo depth data (e.g., Orbbec, Intel RealSense, ZED) for 3D object localization, 6-DoF pose estimation, and scene reconstruction.
• Lift 2D detections and masks into 3D — point clouds, semantic/occupancy maps, and object-centric representations that drive grasping, manipulation, and navigation.
• Own depth data quality: intrinsic/extrinsic and hand-eye calibration, RGB–depth alignment, noise filtering and hole filling, and multi-camera time synchronization.
Edge Deployment & GPU/NPU Optimization
• Deploy vision and VLM models on onboard robot compute (NVIDIA Jetson and NPU-based accelerators) under strict latency, memory, and power budgets.
• Optimize inference through quantization (INT8/FP8/INT4), pruning, knowledge distillation, operator fusion, and model compilation (TensorRT, ONNX Runtime, vendor NPU SDKs).
• Profile and balance workloads across GPU, NPU, and CPU; design multi-model scheduling, batching, and zero-copy pipelines for real-time multi-camera streams.
• Define the edge–cloud split for heavy VLM inference, including fallback behavior when connectivity or compute is constrained.
Research, Data & Integration
• Define evaluation metrics and benchmarks for perception accuracy, latency, throughput, robustness, and safety in real-world deployments.
• Build scalable data pipelines — collection, foundation-model auto-labeling, active learning, and synthetic data — in collaboration with the VinDynamics Data Factory.
• Collaborate closely with Agentic AI, Conversational AI, Manipulation, Navigation, and System Engineering teams to integrate perception into the robot’s perception–reasoning–action loop (ROS 2).
• Translate research breakthroughs into production systems, publish high-impact work, and mentor engineers on the vision team.
Education & Background
• Master’s or PhD in Computer Vision, Artificial Intelligence, Machine Learning, Robotics, or related fields.
• 5+ years of hands-on experience in computer vision, including recent work with VLMs or vision foundation models.
• Proven track record of shipping vision models to production or onto real hardware (robots, edge devices, cameras).
• Proficiency in Python (PyTorch required) and solid C++ for performance-critical deployment.
• Strong system-thinking mindset and rigorous experimental design capability.
Core Expertise
• Vision-Language Models: VLM architectures and fine-tuning (e.g., CLIP/SigLIP encoders, LLaVA, Qwen-VL, InternVL, PaliGemma families); visual grounding, instruction tuning, and hallucination mitigation in visual reasoning.
• Open-vocabulary detection & segmentation: Grounding DINO, OWL-ViT/OWLv2, YOLO-World, Florence-2, the Segment Anything (SAM) family; zero-/few-shot adaptation and multi-object tracking.
• Depth & 3D perception: RGB-D and stereo processing, point clouds (Open3D/PCL), camera calibration, 6-DoF pose estimation, 3D detection/segmentation, and depth-aware multi-sensor fusion.
• GPU/NPU optimization: TensorRT, ONNX/ONNX Runtime, CUDA; PTQ/QAT quantization, pruning, distillation; profiling with Nsight or equivalent; deployment on NVIDIA Jetson and NPU edge platforms.
Preferred Qualifications
• Research publications in top-tier venues such as CVPR, ICCV, ECCV, NeurIPS, ICLR, ICRA, RSS, or CoRL.
• Experience with:
– Vision-Language-Action (VLA) models or embodied AI
– ROS 2 and real-time robot perception pipelines
– Multi-camera rigs and sensor fusion with LiDAR/IMU
– Synthetic data and simulation (e.g., NVIDIA Isaac Sim) for perception training
• Knowledge of:
– Efficient architectures for the edge (lightweight ViTs, token pruning/merging for VLMs)
– Inference and streaming frameworks (TensorRT-LLM, vLLM, Triton, NVIDIA DeepStream)
– Visual SLAM and neural 3D representations (NeRF, 3D Gaussian Splatting)
Personality & Attitude
• Proactive, responsible, and impact-driven with a strong business mindset.
• Hands-on and pragmatic — comfortable debugging on real robots, cameras, and hardware in the lab.
• Excellent communication and cross-functional collaboration skills.
• Strong problem-solving capability and system-level thinking.
• Comfortable working at the frontier of research with evolving constraints.
Attractive income, competitive and commensurate with individual capabilities.
13th-month salary, gifts on public holidays and special occasions, and performance-based bonuses.
Meal allowance, annual company trips, health insurance, and exclusive benefits within the Group’s ecosystem
Clear career development opportunities aligned with the company’s growth, with access to training programs based on capability and job role.
A dynamic and open working environment with diverse cultural and sports activities.
Đang tải…
Nguồn: careers VinDynamics
Vingroup humanoid robotics — Hà Nội
Hanoi · Sync · 125Tăng 1 job so với 3 ngày trước job
Chưa có referrer tại công ty này
Yêu cầu referral vẫn được Admin xử lý thủ công.
| Job | Địa điểm | Thời gian | Hành động |
|---|---|---|---|
Hanoi | Hanoi | ||
Vietnam | Vietnam | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Vietnam | Vietnam | ||
Vietnam | Vietnam | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Vietnam | Vietnam | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Vietnam | Vietnam | ||
Vietnam | Vietnam | ||
Vietnam | Vietnam | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi | ||
Hanoi | Hanoi |
30 job
30 job
Chưa ghi lương
Senior
5+ YoE
Hanoi
Xây dựng kiến trúc thị giác và triển khai mô hình Vision-Language lên robot nhân hình VinDynamics chạy thời gian thực trên edge.
Tóm tắt bằng AI từ JD gốc, đọc JD gốc trước khi ứng tuyển.
At VinDynamics, we design safe, affordable, and intelligent humanoid robots to assist in everyday life — robots for everyone. Backed by Vingroup, Vietnam’s leading technology conglomerate, we are on a mission to make advanced robotics accessible, reliable, and beneficial for billions of people worldwide. By combining cutting-edge AI, world-class engineering, and human-centered design, we aim to seamlessly integrate robots into daily life — enhancing safety, productivity, and happiness at home and beyond.
As our Computer Vision Expert, you will give our humanoid robots the ability to see, understand, and act on the physical world — turning frontier vision-language research into perception that runs in real time on the robot.
Vision-Language Perception & Open-World Understanding
• Design and own the visual perception architecture for humanoid robots, from raw multi-camera input to semantic scene understanding.
• Build and adapt Vision-Language Models (VLMs) for robot-centric tasks: visual question answering, scene description, referring-expression grounding, and instruction-conditioned perception.
• Develop open-vocabulary “detect anything / segment anything” pipelines that detect, segment, and track arbitrary objects in unstructured indoor environments without per-class retraining.
• Fine-tune, adapt, and distill vision foundation models on in-house robot data to improve robustness against lighting changes, occlusion, clutter, and motion blur.
Depth Sensing & 3D Perception
• Process and fuse RGB-D and stereo depth data (e.g., Orbbec, Intel RealSense, ZED) for 3D object localization, 6-DoF pose estimation, and scene reconstruction.
• Lift 2D detections and masks into 3D — point clouds, semantic/occupancy maps, and object-centric representations that drive grasping, manipulation, and navigation.
• Own depth data quality: intrinsic/extrinsic and hand-eye calibration, RGB–depth alignment, noise filtering and hole filling, and multi-camera time synchronization.
Edge Deployment & GPU/NPU Optimization
• Deploy vision and VLM models on onboard robot compute (NVIDIA Jetson and NPU-based accelerators) under strict latency, memory, and power budgets.
• Optimize inference through quantization (INT8/FP8/INT4), pruning, knowledge distillation, operator fusion, and model compilation (TensorRT, ONNX Runtime, vendor NPU SDKs).
• Profile and balance workloads across GPU, NPU, and CPU; design multi-model scheduling, batching, and zero-copy pipelines for real-time multi-camera streams.
• Define the edge–cloud split for heavy VLM inference, including fallback behavior when connectivity or compute is constrained.
Research, Data & Integration
• Define evaluation metrics and benchmarks for perception accuracy, latency, throughput, robustness, and safety in real-world deployments.
• Build scalable data pipelines — collection, foundation-model auto-labeling, active learning, and synthetic data — in collaboration with the VinDynamics Data Factory.
• Collaborate closely with Agentic AI, Conversational AI, Manipulation, Navigation, and System Engineering teams to integrate perception into the robot’s perception–reasoning–action loop (ROS 2).
• Translate research breakthroughs into production systems, publish high-impact work, and mentor engineers on the vision team.
Education & Background
• Master’s or PhD in Computer Vision, Artificial Intelligence, Machine Learning, Robotics, or related fields.
• 5+ years of hands-on experience in computer vision, including recent work with VLMs or vision foundation models.
• Proven track record of shipping vision models to production or onto real hardware (robots, edge devices, cameras).
• Proficiency in Python (PyTorch required) and solid C++ for performance-critical deployment.
• Strong system-thinking mindset and rigorous experimental design capability.
Core Expertise
• Vision-Language Models: VLM architectures and fine-tuning (e.g., CLIP/SigLIP encoders, LLaVA, Qwen-VL, InternVL, PaliGemma families); visual grounding, instruction tuning, and hallucination mitigation in visual reasoning.
• Open-vocabulary detection & segmentation: Grounding DINO, OWL-ViT/OWLv2, YOLO-World, Florence-2, the Segment Anything (SAM) family; zero-/few-shot adaptation and multi-object tracking.
• Depth & 3D perception: RGB-D and stereo processing, point clouds (Open3D/PCL), camera calibration, 6-DoF pose estimation, 3D detection/segmentation, and depth-aware multi-sensor fusion.
• GPU/NPU optimization: TensorRT, ONNX/ONNX Runtime, CUDA; PTQ/QAT quantization, pruning, distillation; profiling with Nsight or equivalent; deployment on NVIDIA Jetson and NPU edge platforms.
Preferred Qualifications
• Research publications in top-tier venues such as CVPR, ICCV, ECCV, NeurIPS, ICLR, ICRA, RSS, or CoRL.
• Experience with:
– Vision-Language-Action (VLA) models or embodied AI
– ROS 2 and real-time robot perception pipelines
– Multi-camera rigs and sensor fusion with LiDAR/IMU
– Synthetic data and simulation (e.g., NVIDIA Isaac Sim) for perception training
• Knowledge of:
– Efficient architectures for the edge (lightweight ViTs, token pruning/merging for VLMs)
– Inference and streaming frameworks (TensorRT-LLM, vLLM, Triton, NVIDIA DeepStream)
– Visual SLAM and neural 3D representations (NeRF, 3D Gaussian Splatting)
Personality & Attitude
• Proactive, responsible, and impact-driven with a strong business mindset.
• Hands-on and pragmatic — comfortable debugging on real robots, cameras, and hardware in the lab.
• Excellent communication and cross-functional collaboration skills.
• Strong problem-solving capability and system-level thinking.
• Comfortable working at the frontier of research with evolving constraints.
Attractive income, competitive and commensurate with individual capabilities.
13th-month salary, gifts on public holidays and special occasions, and performance-based bonuses.
Meal allowance, annual company trips, health insurance, and exclusive benefits within the Group’s ecosystem
Clear career development opportunities aligned with the company’s growth, with access to training programs based on capability and job role.
A dynamic and open working environment with diverse cultural and sports activities.
Đang tải…
Nguồn: careers VinDynamics