PYTHON · NUMPY

Does it obey?

A VLA learned to 'pick the red block' from demos. You say 'pick the BLUE block,' but blue never appeared in training. It will most likely…

Prove the policy follows the instruction rather than chasing a fixed target, by flipping which colour you ask for.

01Challenge

Show that each instruction produces an action pointing at its own named target, the test of real language grounding.

Predict firstA VLA learned to 'pick the red block' from demos. You say 'pick the BLUE block,' but blue never appeared in training. It will most likely…

02Model

A policy that always reaches the same block would look like it understands language while ignoring it. the test is to flip the instruction and check the behaviour flips too. Ask for green and the action points at green; ask for red and it points at red, same scene, different word, different move. This is exactly the measurement behind the headline number on our released nano-VLAs: reach the correct target 97 percent of the time, and 0 percent when the instruction is flipped. That zero is the proof the language is load-bearing, not decoration.

03Guided practice

1. Rebuild the VLA. Each lesson stands alone, so set the policy up again. 2. Flip the instruction. Feed 'green', then 'red', on the same scene. 3. Read the result. Each action aligns with its own target, and the two actions differ. If flipping the word flips the target, the language is doing the work.

04Feedback

PASS WHEN PASS when each instruction points at its own named target and the two actions differ.

If your run fails, check:

A cell errored, run them in order.
Each instruction should aim at the target it names.

05Retrieve & space

What does flipping the instruction test? (whether the policy obeys language or ignores it).
On our released nano-VLAs, what is the success rate when the instruction is flipped? (0%, it reaches the other target).
Why is that zero a good thing? (it proves the language is load-bearing).

06Mastery & project

You can prove a policy obeys language rather than memorizing one target.

The check that separates real grounding from a lucky-looking policy: flip the word, watch the target flip.