Modern electric mobility systems – from Formula E racing to autonomous delivery fleets – operate in environments where success depends on managing multiple competing variables simultaneously. Battery energy, terrain conditions, machine health, weather fluctuations, and traffic congestion all interact in complex ways that are difficult to model and predict.
There is a clear need for efficient modelling of competitive mobility systems, where vehicles, drones, and riders must survive and perform under dynamic, resource-limited, and unpredictable conditions. A unified simulation platform should allow these core parameters - and any future ones - to be added, customized, and evolved as mobility technology and research needs grow.
We propose Drift, a multi-agent simulation platform inspired by the unpredictability and strategic depth of Formula E. We plan to build an environment where autonomous and user-controlled agents compete for efficiency and survival under dynamic, resource-limited conditions.
Key features we plan to build:
- Multi-Agent Arena: A shared ecosystem with each agent responding continuously to changing parameters like battery state, traction, mechanical health, and temperature.
- Real-time Event Dashboard allows admins to dynamically alter the environment by injecting obstacles, and terrain or weather changes to evaluate performance of agents.
- Gameplay will evaluate agents that will be user controlled as well as AI-powered using reinforcement learning.
- Short, iterative rounds (timed or lap-based), with randomized levels and terrain, supporting easy experimentation and evolution.
- Live Visualization and Scoring: 2D/3D rendering of race dynamics, decision highlights, and leaderboards ranking agents by efficiency, adaptability, and survival.
- Fan Interaction Layer: Spectators can give small performance boosts or consumable resources to agents, creating a participatory simulation ecosystem without altering terrain.
tech-stack : We plan to use Rust or Python with FastAPI and WebSockets for the backend, reinforcement learning models using PyTorch for agent behavior, Redis for live state caching, PostgreSQL for telemetry and event logs, and React + Three.js for 3D visualization and interactive spectator control.
Log in or sign up for Devpost to join the conversation.