Release task
Inspect the task files, instructions, source material, and evaluation criteria.
Task requirements
Images load automatically. YAML loads when expanded; audio and other formats remain metadata-only.
Release task
Inspect the task files, instructions, source material, and evaluation criteria.
1 source files available in data/.
Task requirements
I am teaching a class on reinforcement learning and I need to demonstrate how Deep Q-Networks (DQN) work in a visually engaging way. I want to create a browser-based driving simulator that runs 8 parallel environments simultaneously to train a DQN agent. The entire thing should happen in the browser with no server-side dependencies.
The simulator should feature:
The simulation should run at interactive speed — the implementation must be efficient enough that all 8 environments render and train smoothly in real-time without noticeable lag or frame drops.
Use screenshot.png as a reference.
Images load automatically. YAML loads when expanded; audio and other formats remain metadata-only.
12 evaluation criteria
The solution is implemented as a single HTML file (with inline CSS/JS) that runs entirely in the browser with no server-side dependencies.
8 parallel 3D driving environments are rendered simultaneously in a 4x2 grid layout, each showing an independent simulation.
Each environment features a low-poly 3D scene of a car driving on a winding forest road, with curves/bends, grass terrain, and trees lining the road, rendered using Three.js or equivalent WebGL library.
A DQN (Deep Q-Network) agent is implemented in the browser, with a neural network that takes a pixelated low-resolution observation (e.g., 8x8) as input and outputs Q-values for discrete driving actions (e.g., left, right, forward).
The pixelated observation of the first environment (Env 0) is displayed in the metrics/info panel, showing the small grid image that the DQN agent actually sees as its input.
The DQN agent trains across all 8 parallel environments simultaneously, collecting experiences and updating the network from batched transitions.
Training metrics are displayed in real-time: action Q-values for the current state, per-environment reward, and the current epsilon (exploration rate) value.
An average episode reward chart is displayed and updates over training batches, showing the agent's learning progress over time.
Clicking a 'Start Training' or equivalent button initiates the DQN training loop; the vehicles begin moving and the agent starts learning visibly across all 8 environments.
Before training starts, the environments are either static or show random/idle behavior, and metrics panels indicate no training data is available yet.
The simulation runs at interactive speed (smooth real-time rendering and training) with an efficient implementation — all 8 environments should update and render without noticeable lag or frame drops.
The agent shows visible learning progress over time: early episodes exhibit random/erratic driving, while later episodes show the vehicle staying on the road more consistently.