Skip to content

Adaptive Navigation OpenEnv

Adaptive Navigation OpenEnv is a partially observable exploration environment designed for training and evaluating LLM agents on multi-step planning tasks.

Code Demo Video Cerebral Valley

Core trick

Partially observable grid where the agent sees only a local window and must plan key-unlock-goal sequences under an energy budget.

Tags

agents

Description

Adaptive Navigation OpenEnv is a partially observable exploration environment designed for training and evaluating LLM agents on multi-step planning tasks. The agent operates in a 2D world where it only sees a small local observation window and must navigate efficiently under energy constraints. The objective is to collect a key, unlock a checkpoint, and reach a goal while reasoning over incomplete information. The environment includes dynamic obstacles, mission tracking, and reward feedback, making it suitable for reinforcement learning and agentic decision-making research. The project demonstrates a full pipeline including an interactive Streamlit environment, OpenEnv deployment on Hugging Face Spaces, and a minimal HF TRL training scaffold in Colab. This environment provides a compact benchmark for studying autonomous exploration, navigation, and planning with language model agents.
harishchaurasia/Meta_OpenEnv_PyTorch_Hack★ 00 forksPythonCreated Mar 7, 2026 · last push Mar 8, 2026

README excerpt

# Adaptive Navigation Environment

A partially observable 2D grid-world environment designed for training and evaluating LLM / RL agents on **long-horizon planning** and **world modeling**.

Built for the **Meta OpenEnv PyTorch Hackathon** -- fully compatible with [OpenEnv 0.2.1](https://github.com/meta-pytorch/OpenEnv).

---

## Problem Framing

Real-world navigation requires agents to:

1. **Explore** an unknown environment under partial observability.
2. **Plan** multi-step sequences (find key -> unlock door -> reach goal).
3. **Adapt** when the world changes mid-episode (corridors open/close).
4. **Manage resources** (limited energy budget).

This environment distils those challenges into a clean, configurable grid-world that is simple enough to iterate on quickly but rich enough to expose genuine planning failures.

---

## Environment Mechanics

| Element | Symbol | Description |
|---------|--------|-------------|
| Wall | `#` | Impassable terrain |
| Empty | `.` | Walkable cell |
| Agent | `A` | The player |
| Key | `K` | Must be collected to unlock the door |
| Door | `D` | Blocks the path to the goal until unlocked |
| Goal | `G` | Episode ends successfully when reached |
| Hazard | `X` | Penalty on contact (optional) |

The grid is generated with a randomized DFS maze algorithm, then key, door, and goal are placed with BFS-validated reachability guarantees.

---

## Action Space

| Index | Action | Effect |
|-------|--------|--------|
| 0 | Up | Move one cell up |
|

Team