Get to know the role
As a Senior Perception & Prediction Engineer, you will build and ship systems that turn raw multi-sensor data into a reliable, real-time and predictive representation of the world. You will work across a robust modular perception stack and modern learning-based approaches, contributing through model development, evaluation, debugging, integration and on-robot validation.
On top of a shipped multi-sensor detection baseline, you will deepen three connected capabilities: multi-object tracking, generic / open-set world understanding, and motion prediction. You will also help evaluate and productionise end-to-end temporal perception-prediction models and VLA / embodied foundation models for open-vocabulary understanding, long-tail reasoning and task-conditioned robot intelligence.
We are pragmatic about new research: a model earns its place through measurable closed-loop value, reliable grounding, real-time performance and safety. Classical geometry, filtering and modular components remain important as interpretable baselines, safety fallbacks and guardrails. This role offers the opportunity to take promising research from prototype to fleet data, embedded deployment and real-world robot behaviour.
You will report to the Senior Principal Perception & Prediction Engineer and work onsite at a Grab office.
The critical tasks you will perform
You will develop and improve multi-object tracking — data association, Bayesian state estimation (Kalman / EKF / UKF and motion models), track lifecycle, and ID stability through detector gaps, occlusions and crowded scenes — and evolve it toward learning-based tracking, including learned association, joint detection-and-tracking and query / transformer-based trackers.
You will build and harden generic / open-set world understanding — class-agnostic obstacle detection, occupancy grids / occupancy networks / BEV occupancy, clustering and learned generic-object branches — so the robot safely reacts to rare or unknown objects.
You will build motion prediction for pedestrians, cyclists, vehicles and other agents, including multi-modal trajectory forecasting, interaction-aware prediction, occupancy flow, uncertainty estimation and well-defined interfaces into behaviour and planning.
You will develop and evaluate end-to-end temporal perception-prediction approaches, such as joint detection-tracking-forecasting, streaming BEV representations, agent / trajectory queries and learned world models, while maintaining strong modular baselines and production fallbacks.
You will explore VLM / VLA and embodied foundation models for open-vocabulary perception, semantic scene reasoning, task-conditioned understanding, long-tail discovery, auto-labeling and teacher-student supervision. You will test grounding, hallucination, temporal consistency, latency and robustness, then distil, quantise or adapt useful capabilities for on-robot deployment rather than stopping at demos.
You will strengthen multi-sensor and temporal integration across camera, LiDAR, radar and IMU, ensuring geometry, timing, identities, predictions and semantic context remain consistent over time.
You will define quality gates across tracking (MOTA / MOTP-style, ID-switch, fragmentation), prediction (ADE / FDE, miss rate, NLL / calibration), generic-object / occupancy / occupancy-flow recall, closed-loop safety outcomes and embedded latency budgets, and drive down regressions with fleet-log root-cause analysis.
You will partner with Detection, Prediction, Planning, Data & Infra, and Integration teams to move classical, learning-based and foundation-model capabilities from research into reproducible training, replay, simulation, NVIDIA Orin / TensorRT deployment and production-like robot testing.
What Essential Skills You Will Need
At least 3 years hands-on experience shipping or operating perception and / or prediction systems for robotics, AV or ADAS, with strong autonomous-driving experience and a record of improving real-world performance.
Rich experience across classical / geometric and learning-based approaches, with sound judgement on when to use modular methods, end-to-end models or hybrid designs, and how to retain interpretable safety fallbacks.
Proven depth in multi-object tracking — data association, Bayesian filtering, track lifecycle, multi-sensor / temporal association, ID stability and low-latency evaluation; experience with learned or query-based tracking is valuable.
Hands-on experience with generic / open-set perception — class-agnostic obstacle detection, occupancy representations, LiDAR clustering, anomaly / unknown-object handling or fusion with learned detections.
Hands-on depth in at least one modern temporal area: motion forecasting, occupancy flow, streaming / temporal BEV, learned world models, or joint detection-tracking-prediction. You should understand multi-modal uncertainty and the interface between perception, prediction and planning.
Fundamentals in geometry and multi-sensor systems, including practical calibration know-how and the ability to diagnose alignment or time-synchronisation issues that impact model quality.
Proficiency in C++ and Python, with the ability to write production-quality code, build training / evaluation tools and reason about real-time embedded performance.
Experience navigating noisy production-like environments through reproducible experiments, clear metrics, log-driven root-cause analysis, simulation / replay and validation-minded execution.
Self-motivated and dependable, with genuine enthusiasm for autonomous driving, robotics and embodied AI, and the judgement to separate robust progress from research hype.
Demonstrated proficiency in leveraging AI tools and systems to accelerate engineering and research without compromising quality, security or reproducibility.
Good to have
End-to-end perception and prediction: joint detection-tracking-forecasting, occupancy / occupancy-flow prediction, streaming BEV, agent / trajectory queries or world models.
Motion prediction for autonomous systems: multi-modal trajectories, interaction-aware forecasting, uncertainty calibration, planning interfaces, closed-loop simulation or behaviour evaluation.
Experience with VLM / VLA / embodied foundation models, including vision-language grounding, open-vocabulary perception, task or instruction conditioning, semantic scene reasoning, long-tail mining or action-relevant representations.
Foundation-model adaptation and deployment: supervised fine-tuning / LoRA, preference or imitation data, distillation, quantisation, teacher-student pipelines, edge inference and systematic evaluation of hallucination and grounding.
3D / BEV detection and sensor fusion (for example BEVFusion / LSS), temporal models, 4D data and multi-camera / LiDAR / radar integration.
Experience with common robotics / perception libraries and platforms such as OpenCV, PCL, Ceres, ROS 2, NVIDIA Orin, TensorRT or DeepStream, plus strong visualisation and debugging practices.
Familiarity with the model lifecycle: dataset curation, auto-labeling, temporal annotation, active learning, regression detection, inference optimisation and fleet-data feedback loops.
Life at Grab
We care about your well-being at Grab, here are some of the global benefits we offer:
We have your back with Term Life Insurance and comprehensive Medical Insurance.
With GrabFlex, create a benefits package that suits your needs and aspirations.
Celebrate moments that matter in life with loved ones through Parental and Birthday leave, and give back to your communities through Love-all-Serve-all (LASA) volunteering leave
We have a confidential Grabber Assistance Programme to guide and uplift you and your loved ones through life's challenges.
What We Stand For at Grab
We are committed to building an inclusive and equitable workplace that enables diverse Grabbers to grow and perform at their best. As an equal opportunity employer, we consider all candidates fairly and equally regardless of nationality, ethnicity, religion, age, gender identity, sexual orientation, family commitments, physical and mental impairments or disabilities, and other attributes that make them unique.
