Our Technology
AD/ADAS technology is not just tightly integrated vehicles, well balanced sensing suites, and capable SW running on optimized HW, but it also includes a data pipeline that supports scaling and maintenance.
The AD Stack
Our Autonomous Driving stack covers perception, planning, and control as single steps and as a whole. We develop and own the full pipeline from raw sensor data to driving decisions.
Sensing
Production ready sensing suite with 7 long-range cameras, 4 surround cameras, and 5 radars that can be reconfigured into smaller subsets. Hardware, drivers, and sensor integration are developed in-house for a super efficient low level fusion of pixels and radar returns.
Perception
Fully AI-based detection, tracking, and scene understanding across vision and radar. Lidar is optional, currently not used in our configuration. Configurable camera setups (11V, 6V, 5V) are fused with radar to deliver robust perception in complex urban and highway environments.
Planning & Control
AI-based path planning and behavior decision generate safe driving actions. When available, the stack uses a topological map — no HD maps required — and is designed to be E2E-ready as models and compute evolve.
Our road ahead
The stack is E2E-ready today. Our next milestone is full end-to-end training, and a VLM as a slow stack in parallel with the current fast stack to disambiguate complex situations, based on context.
Data pipeline
Our toolchain is not just a collection of tools, but a methodology to scale up. Diversified, well balanced, proprietary datasets together with accurate Ground Truth enable the data pipeline flywheel.
Amba-DAQ
Amba-DAQ is a platform for data recording vehicles that can be assembled and deployed geographically in about five days. A simple interface lets a single operator monitor and run the system, capturing data with the full sensing suite plus ground-truth sensors at roughly 5 TB for eight hours of operation (160 MB/s).
Amba-CAT
Amba-CAT is an auto-annotation tool that generates 3D labels from single or multiple recordings. Built on datasets acquired with a ground-truth-friendly sensing suite (Amba-DAQ), it associates objects across the video sequence and reprojects them onto each image — producing classification, positions, 3D bounding boxes with keypoints, instance segmentation, speed, and trajectory — and enabling low-level fusion training across sensors.
Amba-Lens
Amba-Lens is an AI-based system that analyzes the recorded data; it computes ODD coverage statistics, and provides direct feedback to the acquisition vehicles for targeted data acquisition campaigns to fill possible ODD gaps. It is also a search tool to surface specific scenarios for AD stack testing and helps assemble correctly balanced datasets for training.
Global data acquisition
We are committed to scaling data collection worldwide — deploying Amba-DAQ fleets across geographies to capture diverse ODDs, climates, and traffic cultures, and continuously feed the annotation and training flywheel.
Autonomous Driving in Action
Watch our AD technology on real-world systems.
Built on CV3-AD
Our AD stack runs on Ambarella's CV3-AD SoC family — 5nm ASIL silicon designed for autonomous driving from L2 to L4.
Scalable SOC Family
SOC family supports L2 to L4
Unified SDK
Allows to port the SW stack development across the family
Power Efficiency
High AI performance with minimal power budget; BOM cost savings for ECU/Thermal
Modern GenAI Support
BEV / Transformer for latest AD trends; E2E AI and LLM / GenAI models
Balanced Architecture
Sustainable AI performance even with all functions enabled; patented CVflow® IP with low memory & low power
Leading Image Quality
Long-time image leader in camera market, Ambarella also provides AI-based ISP for automotive image pipeline
Autonomous Driving Software
Fully AI-based perception and planning — E2E ready, no HD maps required.
Sensing technology
Different vision setups (11V, 6V, 5V) fused with radar for robust multi-sensor perception.
Topological maps
Preloaded topological map information — no HD maps required for navigation.
Fully AI-based
Perception and planning driven by AI models end to end across the stack.
E2E ready
Architecture prepared for end-to-end training, with a VLM as slow stack alongside the fast path.
Development toolchain
Not just a collection of tools, but a methodology to scale up.
Proprietary datasets
Diversified, well balanced data with accurate Ground Truth — essential for scalable AI development.
Effective toolchain
Tools and processes designed to be both effective and efficient at production scale.
Data recording at scale
Easy acquisition logistics so fleets can capture data, with our Amba-DAQ platform, wherever the ODD demands it.
Auto-annotation
Extensive dataset labeling with limited human intervention via Amba-CAT.
Data selection
Targeted selection, through Amba-Lens, of scenarios and sequences to close ODD gaps and focus training.