CondenseVLA
A multi-frame vision-language-action (VLA) policy model that introduces a learnable history compression module on the decoder side, compressing historical frame features into a fixed number of slot tokens through learnable queries and cross-attention to enable efficient multi-frame inference.
Timeline (last 90 days)
CondenseVLA is a multi-frame vision-language-action (VLA) policy whose key component is a learnable history compression module on the decoder side.
CondenseVLA uses a set of learnable queries and multi-head cross-attention to distill the observation history into a fixed number of slot tokens.
CondenseVLA's compression occurs at the decoding stage rather than the encoding stage, so the VLM backbone does not need to recompute for every frame.
The number of slot tokens in CondenseVLA is a preset hyperparameter that does not vary with the length of the history window, so decoder computation remains constant.
CondenseVLA introduces auxiliary regularization: it penalizes overlap between slot attention maps and decorrelates slot embeddings.
CondenseVLA raises the success rate on the Simpler-WidowX benchmark from 58.7% to 69.8%, an absolute improvement of 11.1 percentage points.
CondenseVLA achieves a 97.2% success rate on the LIBERO benchmark.
CondenseVLA achieves a 52.8% success rate on real-world tasks.
CondenseVLA achieves a 21.6% success rate on the memory-dependent RMBench benchmark.
The CondenseVLA project page is publicly available on GitHub at https://github.com/FeiyangHong/CondenseVLA.
3 more timeline events
All Facts (13)
CondenseVLA is a multi-frame vision-language-action (VLA) policy whose key component is a learnable history compression module on the decoder side.
50%UnverifiedCondenseVLA uses a set of learnable queries and multi-head cross-attention to distill the observation history into a fixed number of slot tokens.
50%UnverifiedCondenseVLA's compression occurs at the decoding stage rather than the encoding stage, so the VLM backbone does not need to recompute for every frame.
50%UnverifiedThe number of slot tokens in CondenseVLA is a preset hyperparameter that does not vary with the length of the history window, so decoder computation remains constant.
50%UnverifiedCondenseVLA introduces auxiliary regularization: it penalizes overlap between slot attention maps and decorrelates slot embeddings.
50%UnverifiedCondenseVLA raises the success rate on the Simpler-WidowX benchmark from 58.7% to 69.8%, an absolute improvement of 11.1 percentage points.
50%UnverifiedCondenseVLA achieves a 97.2% success rate on the LIBERO benchmark.
50%UnverifiedCondenseVLA achieves a 52.8% success rate on real-world tasks.
50%UnverifiedCondenseVLA achieves a 21.6% success rate on the memory-dependent RMBench benchmark.
50%UnverifiedThe 21.6% success rate on RMBench indicates that tasks relying on long-term memory or reasoning over multiple observations remain a weak point for CondenseVLA.
50%UnverifiedIntroducing hierarchical or adaptive window mechanisms into the compression module may further break through the bottleneck on memory-dependent tasks.
50%UnverifiedThe strategy of a limited window plus compression still hits a ceiling on tasks that require spanning longer time horizons or more complex combinations of observations.
50%UnverifiedThe CondenseVLA project page is publicly available on GitHub at https://github.com/FeiyangHong/CondenseVLA.
50%