Mujoco Robotics Benchmark Design
Budget: €30 – €250 EUR
I want to commission a benchmark problem in MuJoCo that deliberately stretches an AI agent’s ability to adapt to brand-new situations. The aim is not simply to trip the agent up, but to expose genuine gaps so future research can close them. Concretely, the task must centre on complex task sequences—multi-stage interactions where each stage subtly alters the dynamics or goals. When current reinforcement-learning agents attempt to discover a control policy, they should fail for clear, technical reasons rather than by chance or because the environment is poorly specified.
You will work within a strict authoring framework I will provide. It defines naming, observation spaces, termination criteria, episode length, reward signal limits and the minimum reproducibility checks. The scenario itself must remain unambiguous: every object, joint limit, sensor reading and success metric needs to be explicit so that two independent teams will reproduce identical results.
I expect three concrete deliverables:
• A fully-described MuJoCo XML model and any external assets.
• A reference Python wrapper (Gym-style) that exposes the environment exactly as specified, plus a short evaluation script that logs reward curves.
• A brief report (max 2 pages) documenting design choices, why standard agents fail, and how the task targets adaptability to new scenarios.
Code must run on MuJoCo ≥2.3 and Python 3.10. Please keep external dependencies minimal (NumPy, mujoco-python, gymnasium are fine).
If you are excited by pushing the frontier of adaptable robotics control, let’s talk.
You will work within a strict authoring framework I will provide. It defines naming, observation spaces, termination criteria, episode length, reward signal limits and the minimum reproducibility checks. The scenario itself must remain unambiguous: every object, joint limit, sensor reading and success metric needs to be explicit so that two independent teams will reproduce identical results.
I expect three concrete deliverables:
• A fully-described MuJoCo XML model and any external assets.
• A reference Python wrapper (Gym-style) that exposes the environment exactly as specified, plus a short evaluation script that logs reward curves.
• A brief report (max 2 pages) documenting design choices, why standard agents fail, and how the task targets adaptability to new scenarios.
Code must run on MuJoCo ≥2.3 and Python 3.10. Please keep external dependencies minimal (NumPy, mujoco-python, gymnasium are fine).
If you are excited by pushing the frontier of adaptable robotics control, let’s talk.