Editing AI for Robotics (section)

== <span style="color: #FFFFFF;">Analyzing</span> ==
{| class="wikitable"
|+ Robot Learning Approach Comparison
! Approach !! Data Needed !! Generalization !! Real-World Robustness !! Training Cost
|-
| Hand-programmed || None (code) || None (rigid) || High in structured env || High (engineering time)
|-
| Imitation learning || Demos (hours) || Low || Moderate || Low-medium
|-
| RL in sim (domain rand.) || Millions of sim steps || Medium || Moderate (sim-to-real gap) || High (GPU cluster)
|-
| RL in real world || Millions of real interactions || Medium || High || Extreme (robot wear)
|-
| Foundation model policy || Large pre-training dataset || High || Moderate || Very high (pre-training)
|}

'''Failure modes and challenges:'''
* '''Sim-to-real gap''' — Physical parameters (friction, joint stiffness, sensor noise) differ between simulation and reality. Policy trained in simulation fails on the real robot. Domain randomization and real-world fine-tuning mitigate this.
* '''Contact dynamics modeling''' — Rigid body simulation is inaccurate for soft objects, deformable materials, and complex contact (cloth, liquids, granular material). Manipulation of everyday objects remains challenging.
* '''Out-of-distribution generalization''' — Robot policies trained on specific objects fail on similar but novel objects. Foundation model policies improve this but don't eliminate it.
* '''Safe exploration''' — An RL agent exploring to learn may damage itself, nearby humans, or the environment. Safe RL and constrained optimization are active research areas.
* '''Long-horizon task composition''' — Combining multiple primitive skills into complex multi-step tasks (cook a meal, assemble furniture) remains very difficult without hierarchical planning.
</div>

<div style="background-color: #483D8B; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">