BLOG ILLUSTRATION / EXECUTION STORY

Context adaptation or policy-weight updates

The rolling context receives a demonstration and fresh observations. Pretrained weights do not change during this operation. A separate task dataset supplies input for gradient tuning. The reported conditions are static annotations, not a measured improvement trajectory.

Context adaptation or policy-weight updatesThe rolling context receives a demonstration and fresh observations. Pretrained weights do not change during this operation. A separate task dataset supplies input for gradient tuning. The reported conditions are static annotations, not a measured improvement trajectory. CONTEXT · FIXED WEIGHTS WEIGHTS · SEPARATE TRAINING DATA 3–12 s demo video + actions 30 s context prompt + live state GEN-1.5 policy physical pretraining Robot execution 100 Hz actions 59% ±10% reported Task data ~5 min / ~50 demos Gradient tuning 10 reported steps Updated policy 83% ±9% reported Observations update context.
Figure 10. GEN-1.5 uses a 3–12 second sensorimotor demonstration in a 30-second rolling context, and the developer’s report gives a 100 Hz rate for action trajectories. The developer’s report gives 59% ±10% without gradient steps for ten simple short-horizon tasks. The 83% ±9% result uses approximately five minutes of task data and ten gradient steps. The rates are not comparable across benchmarks. The author made this diagram. Context updates do not change weights, and tuning is a separate condition. Open full-size SVG