This article is based on announcements made around AWS re:Invent 2019. Original reporting credit: Analytics India Magazine.
What Is AWS DeepRacer Evo?
AWS DeepRacer is Amazon’s 1/18th-scale autonomous racing car designed as a hands-on platform for learning reinforcement learning (RL). First announced at AWS re:Invent 2018, it lets machine learning developers train RL models in a simulator and then deploy them to a physical car that races on a real track.
At AWS re:Invent 2019, Amazon announced AWS DeepRacer Evo, a hardware and software upgrade to the original platform. The upgrade introduces new sensors, additional race modes, an expanded global league, and more configurable developer tooling — all aimed at making the competition more technically demanding and accessible to a wider audience.
Hardware Upgrades: Stereo Camera and LIDAR
The most significant physical change in DeepRacer Evo is the addition of two new sensors:
- Stereo camera: Replaces (or supplements) the single front-facing camera in the original DeepRacer. A stereo pair allows depth perception, giving the model spatial information about the environment rather than a flat 2D view.
- LIDAR (Light Detection and Ranging): A laser-based sensor that measures distances to surrounding objects by timing reflected light pulses. LIDAR enables the car to detect obstacles in its path and react accordingly.
Together, these sensors unlock a new category of challenge: obstacle detection and avoidance. Where the original DeepRacer focused purely on staying on the track as fast as possible, Evo requires developers to train models that can perceive and react to dynamic track conditions — a significantly harder RL problem.
For existing DeepRacer owners, Amazon announced a sensor upgrade kit that can retrofit the new hardware onto the original car. At the time of announcement, the kit was expected to ship in early 2020.
New Race Modes
The original DeepRacer League used a single format: time trial, where each car races alone on the track and the fastest lap wins. DeepRacer Evo introduces two additional competitive modes:
Object Detection & Avoidance
Static obstacles are placed on the track, and the car must navigate around them without leaving the track boundary. This mode directly leverages the LIDAR sensor and forces developers to incorporate obstacle-awareness into their reward functions — something that wasn’t required in time trial racing.
Head-to-Head Racing
Two DeepRacer cars compete simultaneously on the same track. The opposing car acts as a dynamic obstacle, meaning the model must handle a moving object with unpredictable behavior. This mode requires a substantially more sophisticated policy than time trial racing because the optimal racing line changes based on the position of the competitor.
Expanded DeepRacer League
Amazon extended the DeepRacer League to include:
- Eight new physical race events spanning five additional countries, piggybacking on the AWS Summit event series.
- 18 virtual tracks available in the online racing simulator, giving developers a wider variety of track geometries to train and test against.
- Community-hosted virtual races, allowing any developer or organization to create and run their own private or public racing events through the AWS console.
The community race feature is particularly notable for enterprise ML teams. Organizations can run internal DeepRacer competitions as a structured ML learning exercise — something companies like Accenture had already been doing at scale in partnership with AWS before the Evo announcement.
Console and Model Training Changes
Beyond the hardware, Amazon updated the AWS DeepRacer console — the browser-based interface where developers define, train, and evaluate their RL models. Key changes include:
- Additional hyperparameters: More training parameters are exposed for tuning, giving experienced practitioners finer control over the RL training process (e.g., learning rate schedules, discount factors, entropy settings).
- More complex reward function inputs: The reward function — the Python function developers write to tell the model what constitutes good behavior — receives additional input variables, particularly around sensor data from the stereo camera and LIDAR.
- Choice of neural network topology: Developers can now select from multiple network architectures for their policy model, rather than being locked into a single default structure. This allows experimentation with deeper or differently shaped networks suited to more complex observation spaces.
These changes make the model design process more complex, but that complexity is the point. As the original league matured, leading competitors converged on similar hyperparameter configurations, reducing the diversity of strategies. The expanded configurability raises the ceiling for what developers can explore.
Why the Original League Was Getting Predictable
In the time trial format with a fixed sensor configuration, the optimization problem is relatively narrow: find reward function parameters and hyperparameter values that produce a fast, stable lap. Once a developer found a well-performing configuration, it could be replicated with minor adjustments across subsequent league events.
DeepRacer Evo increases the number of variables a developer must optimize simultaneously — sensor fusion, obstacle response, head-to-head strategy, and network architecture — which makes it much harder to copy a single winning configuration from one race to the next. This should produce a more technically varied and competitive field.
What This Means for ML Developers
AWS DeepRacer has always been positioned as a learning tool first and a competition second. The Evo upgrade pushes developers toward reinforcement learning concepts that are directly relevant to real-world autonomous systems:
- Sensor fusion: Combining camera and LIDAR data requires thinking carefully about observation spaces and how different data types are processed by the policy network.
- Multi-agent RL: Head-to-head racing is a simplified introduction to competitive multi-agent environments, a research area with significant industry applications.
- Reward shaping under uncertainty: Writing reward functions that produce good behavior in the presence of dynamic obstacles is considerably harder than rewarding simple track-following.
For teams looking to build ML competency, the expanded community race feature means DeepRacer can now function as an internal learning and evaluation program, not just a route to the public league.
Summary
| Feature | Original DeepRacer | DeepRacer Evo |
|---|---|---|
| Camera | Single front-facing | Stereo camera |
| Proximity sensing | None | LIDAR |
| Race modes | Time trial only | Time trial, Object Avoidance, Head-to-Head |
| Virtual tracks | Limited set | 18 tracks |
| Community races | Not available | Supported via console |
| Network topology choice | Fixed | Selectable |
AWS DeepRacer Evo is a meaningful step forward for Amazon’s ML education platform. The hardware additions are significant enough to change the nature of the RL problem developers must solve, and the community race feature removes the barrier to running private league events. Whether you are new to reinforcement learning or looking to deepen existing skills, the 2020 DeepRacer season offers a more challenging and technically rich environment than its predecessor.