RoboMamba-DINO-LLaMA-ACT (70k Steps / 62.4 Epochs Multi-Task Policy)
Multi-task imitation learning policy combining:
- Vision Backbone: Pre-trained DINOv2 ()
- Language Backbone: TinyLlama-1.1B ()
- Sequence Model: RoboMamba (Bidirectional Mamba Selective State Space Model, {model}=768$)
- Action Decoder: Action Chunking with Transformers (ACT) + CVAE (=16$, =8$)
Training Details:
- Dataset: 110 Franka manipulation episodes (100 DROID episodes + 10 competition pick demos, 35,878 frames) across 57 manipulation tasks.
- Training Duration: 70,000 optimizer steps (~62.4 epochs).
- Convergence: L1 action reconstruction loss dropped from 0.6418 to 0.0631.
- Docker Image: