Summary
This paper introduces DTR (Drop-Then-Recovery), a protocol that systematically removes blocks from a pretrained VLA model and fine-tunes the result to measure which components are genuinely necessary. It reveals a striking asymmetry: language backbones are highly redundant for standard manipulation tasks, while vision and action pathways are far less tolerant to removal.
Key Contributions
- DTR protocol: drop blocks, fine-tune, measure post-removal recoverability across architectures
- GateProbe: a one-shot virtual-gate sensitivity metric ranking blocks by contribution to downstream action loss
- Strong empirical finding: removing half of LLM blocks in OpenVLA-OFT actually improves LIBERO success rate from 95.0% to 98.3%; retaining only two language blocks still recovers baseline performance
- Vision and action pathways are substantially less compressible than the language backbone
Significance
Provides actionable guidance for VLA compression and architectural design — language backbone capacity inherited from large VLMs far exceeds what robot manipulation actually needs, opening a direct path to leaner and faster VLA models without sacrificing performance.