Waymo Introduces a New Simulator to Advance Autonomous Vehicle Research
Waymo introduces new tools for autonomous AI researchers
Created on October 23|Last edited on October 23
Comment
Waymo, the autonomous vehicle subsidiary of Alphabet, has recently launched Waymax, a novel simulator designed to advance research in the development of self-driving cars. Unlike traditional simulators, which often come with predefined agents like pedestrians, cyclists, and other cars, Waymax aims to offer a more robust and scalable environment for training intelligent agents.
Background
Drago Anguelov, head of research at Waymo, highlighted that traditional simulators typically come with pre-scripted agents, limiting their behavioral realism. Waymax, however, builds upon large datasets gathered from real-world observations, enhancing the agents' ability to imitate complex behaviors and interactions. This focus on stronger imitative components aims to lead to more reliable autonomous driving systems.
Lightweight But Effective
Despite being termed "lightweight," Waymax is engineered to support complex multi-agent interactions. Instead of intricate visual details, the simulator emphasizes bounding boxes and minimal representations to help researchers concentrate on behavioral complexities rather than visual aesthetics. This design choice allows researchers to iterate rapidly, testing new hypotheses and training agents more effectively. One can think of the simulator as an equivalent of OpenAI Gym, but for driving tasks instead of games.
Features and Capabilities
Waymax is developed using JAX, a decision that ensures compatibility with hardware accelerators and provides various functional transforms. It offers multiple features, including data management capabilities that support the Waymo Open Motion Dataset, either loaded locally or streamed via Google Cloud.
Moreover, the simulator implements several evaluation metrics, such as Log Divergence, Collision, and Offroad indicators, among others. It also supports different models for vehicle simulation, like direct state-based control and control via the kinematic bicycle model.
Advancements in Multi-Agent Environments
One of Waymo's key objectives is to address challenges in multi-agent environments. Last year, the company hosted a challenge called "Simulated Agents," but realized the need for a more robust training environment. Collaborating with Google Research, Waymo has now developed Waymax to fill this gap.
A Resource for the Research Community
While the software is available on GitHub, it is not intended for commercial use. Instead, Waymo aims to provide the academic and research communities with advanced tools to expedite autonomous vehicle research. Waymax aligns with Waymo's larger initiatives like the Waymo Open Dataset, providing valuable resources to researchers around the globe.
Future Directions
Anguelov indicates that Waymo will likely host new challenges using Waymax as the training environment, allowing them to gauge the progress in solving specific problems related to autonomous driving. The simulator is also expected to facilitate improvements in reinforcement learning techniques, thereby contributing to the development of self-driving cars exhibiting more complex and emergent behaviors.
Sources:
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.