Event
CondenseVLA: Learnable History Compression Improves Multi-Frame VLA Policies
1 reports · 1 sources
Feiyang Hong et al. propose CondenseVLA, a multi-frame VLA policy that adds a learnable history compression module at the action decoding end. It uses learnable queries and multi-head cross-attention to distill the per-frame features cached by the VLM backbone into a fixed number of slot tokens, and applies auxiliary regularization that penalizes overlap in slot attention maps and decorrelates slot embeddings to reduce redundancy. The paper has been accepted to NeurIPS 2026. On the evaluated tasks, CondenseVLA improves the Simpler-WidowX success rate from 58.7% to 69.8% (+11.1 percentage points), reaches 97.2% on LIBERO, and reaches 52.8% on real-world tasks; on the memory-dependent RMBench tasks, the success rate is 21.6%.
Report timeline
- CondenseVLA: Improving Multi-Frame VLA Policy Efficiency with Learnable History Compression
CondenseVLA proposes adding a learnable history compression module to the decoding side of VLA policies. It uses learnable queries and multi-head cross-attention to distill the per-frame features cached by the VLM backbone into a fixed number of slot tokens, and applies auxiliary regularization that penalizes overlap among slot attention maps and decorrelates slot embeddings to reduce redundancy.
arXiv cs.RO ·