Tried pollen-robotics/microduck_rl: A Fast RL Sandbox for Embodied AI

pollen-robotics/microduck_rl is gaining attention quickly—+147 GitHub stars today—because it targets a practical gap in robotics development: reusable reinforcement-learning environments for the Microduck platform, built around mjlab.

The project appears focused on making simulation-driven training easier to reproduce and iterate on. Instead of wiring every experiment from scratch, developers can work with a more structured environment for testing locomotion policies, reward functions, and controller behavior before deploying to hardware. That is valuable for indie robotics teams: simulation is cheaper than breaking real robots, and fast experiment loops matter more than having a huge infrastructure budget.

Quick architecture test

A sensible setup is to keep RL training local while routing planning, experiment summaries, or agent-assisted debugging through an OpenAI-compatible gateway:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://b-lost.com/v1",
    api_key=os.environ["B_LOST_API_KEY"],
)

response = client.chat.completions.create(
    model="claude-fable-5",
    messages=[
        {
            "role": "system",
            "content": "You are an RL experiment assistant. Analyze training metrics concisely."
        },
        {
            "role": "user",
            "content": "Summarize this Microduck run and suggest the next reward adjustment."
        },
    ],
)

print(response.choices[0].message.content)

The important separation is architectural: mjlab and policy training stay on your GPU, while the language model handles experiment analysis, configuration generation, and troubleshooting. You can also point Cursor, Cline, Roo Code, Windsurf, Aider, or LibreChat at the same custom base URL.

For repeated system prompts and large metric context, B-Lost’s native Anthropic /v1/messages Prompt Caching can reduce cache-hit input costs by 90%, which is useful when running many evaluation cycles. Its listed pricing is 20% below official rates, though actual usage still depends on the selected model and token volume.

My take: this is worth watching if you are building affordable sim-to-real workflows. The project’s traction suggests Microduck is becoming a practical playground for robotics RL—not just another demo repository.