AAM-GR00T-N1.6 — Action-Aware Adaptive Memory (v1)
This checkpoint adds an Action-Aware Adaptive Memory (AAM) module on top of the HAMLET window architecture for GR00T N1.6, trained on RoboMME. The memory module uses a log-spaced 8-frame window [-1024, -512, -256, -128, -64, -32, -16, 0] env-steps; each past observation slot receives a learned per-slot soft gate weight, feeding a 2-layer MemoryTransformer that conditions the DiT flow-matching action head. A second prediction head performs action-conditioned forward prediction as an auxiliary loss. Trainable components: DiT + projector + moment tokens + top-4 LLM layers + gate + prediction head; the remainder of the VLM backbone is frozen.
Evaluation Results
Eval harness: 10 episodes/task, checkpoint-60000. Vanilla GR00T N1.6 = 10.12 avg, HAMLET baseline = 16.5 avg.
| Task Category | AAM (this ckpt) | HAMLET baseline |
|---|---|---|
| Counting | 5.0 | — |
| Permanence | 20.0 | ~19.5 |
| Reference | 12.5 | — |
| Imitation | 5.0 | — |
| Total avg | 10.6 | 16.5 |
The adaptive gate is strong at selective recall (Permanence ties HAMLET) but weak at temporal integration (Counting).
Training Details
- Base model:
nvidia/GR00T-N1.6-3B - Training dataset: RoboMME (lerobot format)
- Steps: 60,000
- Global batch size: 32
- Precision: bf16