Applications algorithms

The new AI Inference processor
background image

HORIZON AI ENGINEERS WINS WAYMO ALGORITHMS
CHALLENGE IN JUNE 2020

HORIZON AI ENGINEERS WIN WAYMO CHALLENGE
AGAIN IN JUNE 2021

Under constraints of
< 70ms latency and > 70 APH / L2 (accuracy)
Horizon Won
The 2 main World Championships
1. Highest accuracy 3D detection model under
latency constraint (AFDetv2)
2. Most efficient model with lowest latency under
accuracy constraints (AFDetv2-Base)

LIDAR PERCEPTION

LiDAR perception models run very efficiently on Journey BPU and can be fused with multi-camera perception

BIRD’S EYE VIEW (BEV) PERCEPTION ALGORITHMS

A new perception paradigm (available on Journey 5 driving solutions)

Current per-camera perception output in image space has limitations:

- Need rules to transform the representation into 3D space
- Incomplete intermediate results. Could miss important perception when objects stretches across cameras
- Per-sensor processing requires late fusion
- Post processing is not learnable. Corner cases require large engineering effort

BEV: AI models directly output their perception results in 3D space

- Compose a complete set of perception formats with the neural networks
- Sensor stream fusion with AI models aka middle fusion
- End-to-end learnable and data driven improvements (rather than code driven), saves engineering effort
- Challenges: New fusion architecture. Perception task completeness (static/dynamic – Semantic – Structure). 4D ground truth generation

HORIZON ADVANCED LEARNED BASED PLANNING

AI based Planning addresses complex scenario and improves the driving experience

In AD, the Planning stage consumes the BEV perception representation to support automated driving functions. Planning challenge is to analyze many possible outcomes and make best decisions to ensure a safe and comfortable driving, while minimizing compute.

Rule based explicit planning has inherent limitations:

- Many possible options require to restrict ODDs and enact arbitrary rules
- Low data utilization
- Conservative. Not human-like behavior

Imitation AI planning learns from expert.
Driver data Interactive AI planning learns from interaction and road experiences and takes multi-agent prediction into account

Efficient AI Computing

Let’s talk about our journey together!