Sichang Su (Destiny)

PhD Student@UT Austin | Interested in Robot Learning

[Sichang Su (Destiny) @SichangSu42538]

Your fine-tuned VLA does most of a long-horizon task, then gets stuck at a few critical steps.

Don't retrain it. Don't collect more demos of what already works.

Let it practice the part that fails, on the real robot.

PARTS, w/ @ling_feng_sun (project lead) 🧵

1/8

[Sichang Su (Destiny) @SichangSu42538]

Takeaway: you don't need to pretrain, or even retrain, a base policy to make it proficient. Most of the value of real-world data is at the few subtasks that fail. Spend the robot-minutes there, and let agents run the loop.

📄https://t.co/EEKVv7T9I5
🌐https://t.co/XJTfsscESi
8/8

[Sichang Su (Destiny) @SichangSu42538]

Your fine-tuned VLA does most of a long-horizon task, then gets stuck at a few critical steps.

Don't retrain it. Don't collect more demos of what already works.

Let it practice the part that fails, on the real robot.

PARTS, w/ @ling_feng_sun (project lead) 🧵

1/8

[Sichang Su (Destiny) @SichangSu42538]

What "minimal human intervention" means here:

Humans pick bottlenecks, write contracts, reset when needed (e.g. remove seated earbuds), and check earbud verifier labels. No human action corrections or policy switching during RL. LEGO/cable reward labels are automatic.

7/8

[Sichang Su (Destiny) @SichangSu42538]

@celkzowie The website is really helpful for me! Thanks a lot

[Sichang Su (Destiny) @SichangSu42538]

@gio_ramponi Are remote interns welcome?