Building reliable sim agents

Building reliable sim driving agents by scaling self-play

State and observation space

Agents are provided with a partial view of the environment through a relative coordinate frame
We show several state <> observations below

Simulator state

A bird's eye view of a traffic scenario

Agent observation

The same traffic scenario through the "eyes" of the agent

Simulator state

A bird's eye view of a traffic scenario

Agent observation

The same traffic scenario through the "eyes" of the agent

Simulator state

A bird's eye view of a traffic scenario

Agent observation

The same traffic scenario through the "eyes" of the agent

Results 4.1 & 4.2 | Examples

Samples of perfectly solved scenarios

Intersection

Intersection

Three-way intersection

Intersection

Samples of failure modes | Collided

Samples of failure modes | Off-road

Roundabout

Collision with road edge

Collision with road edge

Stop sign on the road

Results 4.3.2 | Fast finetuning

Before finetuning

Agents freeze

After 15 minutes of finetuning

Agents effectively reach goals placed behind them

Page updated

Google Sites

Report abuse