5 hours ago
Shanghai, ChinaStaff+
Responsibilities
- Develop and improve multi-object tracking using data association, Bayesian state estimation, Kalman/EKF/UKF filtering, motion models, track lifecycle management, and learning-based trackers.
- Build generic and open-set world-understanding capabilities including class-agnostic obstacle detection, occupancy grids, occupancy networks, BEV occupancy, clustering, and unknown-object handling.
- Develop motion prediction for multiple agent types using multimodal trajectory forecasting, interaction-aware prediction, occupancy flow, uncertainty estimation, and interfaces to behavior and planning.
- Evaluate and productionize temporal perception-prediction systems, including joint detection-tracking-forecasting, streaming BEV representations, trajectory queries, and learned world models.
- Explore VLM/VLA and embodied foundation models for open-vocabulary perception, scene reasoning, task-conditioned understanding, long-tail discovery, auto-labeling, and teacher-student supervision.
- Integrate camera, LiDAR, radar, and IMU data while maintaining geometric, temporal, identity, prediction, and semantic consistency.
- Define quality gates and investigate regressions using tracking, prediction, occupancy, safety, latency, and fleet-log metrics.
- Partner with Detection, Prediction, Planning, Data & Infra, and Integration teams to move systems into reproducible training, replay, simulation, embedded deployment, and robot testing.
Requirements
- At least 3 years of hands-on experience shipping or operating perception and/or prediction systems for robotics, autonomous vehicles, or ADAS, with strong autonomous-driving experience.
- Rich experience with classical/geometric and learning-based approaches, including judgment about modular, end-to-end, and hybrid designs and interpretable safety fallbacks.
- Proven depth in multi-object tracking, Bayesian filtering, track lifecycle management, multisensor and temporal association, ID stability, and low-latency evaluation.
- Hands-on experience with generic/open-set perception, occupancy representations, LiDAR clustering, anomaly or unknown-object handling, or fusion with learned detections.
- Depth in at least one temporal area such as motion forecasting, occupancy flow, streaming/temporal BEV, learned world models, or joint detection-tracking-prediction.
- Fundamentals in geometry and multisensor systems, including calibration and diagnosis of alignment or time-synchronization issues.
- Proficiency in C++ and Python for production code, training and evaluation tools, and real-time embedded performance work.
- Experience with reproducible experiments, metrics, log-driven root-cause analysis, simulation/replay, and validation-oriented execution.
- Demonstrated ability to use AI tools and systems to accelerate engineering and research while maintaining quality, security, and reproducibility.
- Preferred experience includes end-to-end perception-prediction, multimodal motion forecasting, VLM/VLA or embodied foundation models, foundation-model adaptation and deployment, 3D/BEV detection, sensor fusion, robotics/perception libraries, edge inference, and model lifecycle operations.
Benefits
- Term Life Insurance and comprehensive Medical Insurance.
- GrabFlex benefits package customization.
- Parental leave, Birthday leave, and Love-all-Serve-all volunteering leave.
- Confidential Grabber Assistance Programme.
- FlexWork arrangements with differentiated hours.
- The role is onsite at a Grab office.
Categories
ML EngineeringRobotics
About Grab
Grab builds a Southeast Asia superapp for mobility, deliveries, and financial services, serving consumers, drivers, merchants, and enterprises. It monetizes via ride and delivery commissions, merchant services (such as ads and subscriptions), and payments and digital banking products. Founded in 2012 and headquartered in Singapore, Grab operates across multiple Southeast Asian markets and is publicly listed on NASDAQ.