Editing AI for Robotics (section)

== <span style="color: #FFFFFF;">Evaluating</span> ==
Robot AI evaluation is more complex than pure software evaluation because physical performance matters:

'''Simulation benchmarks''': DeepMind Control Suite (locomotion tasks), Meta-World (manipulation tasks), RoboSuite, OpenAI Gym MuJoCo — measure policy performance in simulation. Fast to evaluate, but simulation scores don't always transfer to real performance.

'''Real-world benchmarks''': ALOHA (bimanual manipulation), FurnitureBench (furniture assembly), RoboAgent (diverse household tasks) — measure real robot performance. Ground truth, but slow, expensive, and hard to standardize.

'''Task completion rate''': What fraction of task trials does the robot complete successfully? Specify separately: pick rate, place success rate, task success rate (all steps completed). Measure across different object sizes, weights, positions, and lighting conditions.

'''Cycle time and throughput''': For industrial robotics, task completion time matters as much as success rate. A slow but reliable robot may be less valuable than a fast, slightly less reliable one.

Expert practitioners quantify '''failure mode distribution''': what types of failures occur and at what rate? Grasp failures (couldn't pick up)? Placement errors? Dropped objects? This guides targeted improvement more than aggregate success rate alone.
</div>

<div style="background-color: #2F4F4F; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">