Zima · 2025-12-09 · 4 hours
Goal: Get a functional / smooth arm control system working for zima
What I did:
- read http://www.incompleteideas.net/IncIdeas/BitterLesson.html (Sutton, 2019) + Era of experience paper
- interesting perspective on the research direction for ai. Interesting how it ties into embodied systems
- Improved error estimation with joint difference error penalty (current_joint_state - next_joint_state)
- Removed kd_tree warm start once a current_joint_state is provided
- visualized arm and compared to real life. Improved arm model to match real world more closely
What worked:
- Got much smoother arm movement after various kinematics system tweaks. Fully integrated with ros2 control system.
- Having a more accurate arm model (tweaking link/joint transformations) improved stability as well
-
What failed:
- Still some slight jitter and inconsistency at extreme arm angles (close to workspace boundary especially with high z), but overall much better than initial versions.
Key learning:
- Modelling accuracy is quite important
- starting point / warm start of gradient descent algorithm is quite important
Next session:
- Begin writing data collection code for more complex demos
- arm + tracks + multi camera data
- finish and test multi camera driver
Time spent: 4 hours