A Gymnasium interface for a real $300 robot — mock backend works now, no hardware needed

I’ve been building an open-source robot platform (ESP32-S3 + Raspberry Pi 5, mecanum wheels, ROS2) with a Python SDK that wraps everything behind a simple API:

from threewe import Robot
async with Robot(backend="mock") as robot:
    await robot.move_to(x=3.0, y=2.0)
    scan = robot.get_lidar_scan()  # 360 rays
    pose = robot.get_pose()

The same code runs on backend="mock" (pure Python, zero deps beyond numpy), backend="gazebo", or backend="real" — no changes needed.

What works today:

• Mock backend with 2D kinematics, collision detection, LiDAR raycasting — 309 tests passing

• gymnasium.make("3we/Navigation-v1") for RL training

• VLM integration (GPT-4o / Qwen-VL → robot actions)

• Data recording in HDF5 format (LeRobot-compatible export planned)

Try it (30 seconds):

git clone https://github.com/telleroutlook/3we-robot-platform.git
cd 3we-robot-platform && pip install -e sdk/threewe/
python examples/navigate_office.py

Requires only Python 3.10+ and numpy.

What I’m looking for feedback on:

1. For RL researchers: is the Gymnasium env interface what you’d expect? What observations/action spaces do you want?

2. For VLM/VLA researchers: the SDK has robot.execute_instruction("go to the red door") — what would make this useful for your workflow?

3. For anyone doing imitation learning: trajectory recording exports to HDF5. Would direct LeRobot dataset format be more useful?

Hardware is <$500 to reproduce (full open BOM + KiCad PCB files in the repo). The full Sim2Real chain validation is the next milestone.

GitHub: https://github.com/telleroutlook/3we-robot-platform

Dev Log (design decisions + honest status): https://3we.org/blog/dev-log-001/