Inspiration

As deep reinforcement learning (RL) research accelerates, standardized agent evaluation remains scattered and time-consuming. We were inspired by OpenAI Gym and DeepMind Lab but wanted a unified, easy-to-deploy system with:

GridWorld environments for early-stage algorithm debugging. Robotic simulation layers for real-world generalization. Our goal: make RL experimentation as reproducible and modular as unit testing in software engineering.

What it does

Devprime ADK (Agent Development Kit) provides:

Benchmarkable GridWorlds with curriculum difficulty

Robotic agent wrapper environments using gym and PyBullet

Evaluation suite with side-by-side plots (reward curves, steps, wall time)

Multi-agent support for cooperative/competitive settings

TorchScript-exportable agents for deployment or analysis

Colab-ready training + visualization notebook

GPU-enabled RL agent trainer with model save/load support

How we built it

Stack:

Python 3.10

PyTorch for agent models

OpenAI Gym for environment interfaces

NumPy for reward shaping + state representation

PyBullet for robotics tasks

Matplotlib + Seaborn for visual benchmarking

TorchScript for model export

Google Colab for GPU support + live demo notebooks 📁 devprime/ ┣ 📂 envs/ ┃ ┣ gridworld.py ┃ ┗ robot_env.py ┣ 📂 agents/ ┃ ┣ dqn.py ┃ ┣ ppo.py ┗ 📂 utils/ ┗ logger.py, visualizer.py

Challenges we ran into

Balancing environment complexity with agent training time

Making GridWorld compatible with multi-agent RL setups

Stable reward curves with robotics physics simulation

Exporting Torch models that work across environments

Keeping training GPU-compatible and Colab-ready

Accomplishments that we're proud of

Smooth benchmarking plots comparing DQN, PPO, and custom models

Robotics simulation agent successfully learns target-reaching

Plug-and-play API for defining new environments or agent types

Trained multi-agent behaviors in adversarial GridWorld

Lightweight enough to run fully in Colab with GPU acceleration

What we learned

Fine-tuning reward shaping matters even in synthetic environments

TorchScript can greatly simplify agent sharing and analysis

Maintaining code modularity across environments pays off when scaling

Debugging multi-agent training requires structured logging and replay

What's next for Devprime

Add Unity ML-Agents interface for high-fidelity physics

Add leaderboard-driven benchmarking with reproducibility scripts

PyPI package + GitHub Action for auto-testing agents

Visual RL agent demo with real-time action viewer

Integration with LLM-based decision agents in hybrid environments

Built With

Share this project:

Updates