Experiment 4 — DAgger Steering

Goal

Test whether a fixed representation-space steering vector trained by DAgger can redirect a frozen robot policy from one prompt semantics to another without modifying the model weights.

Current Read

Seen-scene behavior looks materially better than raw simulator reward suggests, but transfer off the left/right train axis is not yet established.

Sub-Experiments

4.1

DAgger Steering on GR00T N1.6 : Seen-Scene Fit vs Up/Down Holdout

A frozen nvidia/GR00T-N1.6-fractal is steered by a learned image-token delta trained from DAgger relabeling: student prompt = pick up the orange, teacher prompt = pick up the bottle. The published delta is trained only on left/right scenes across multiple student rollouts. Current intermediate result: strong qualitative left/right steering on seen scenes, but no simulator success yet on the new up/down holdout axis in the first preview.

Intermediate result

Status

This experiment page is intentionally intermediate. It is meant to show the DAgger steering protocol, the clean left/right success story, and the first failed holdout transfer results before a final polished write-up.