Unverified50% confidenceFactExact time
CondenseVLA is a multi-frame vision-language-action (VLA) policy whose key component is a learnable history compression module on the decoder side.
1
Sources
50%
Confidence
Long-term
Relevance
10/20/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe construction pipeline of the VG-Pairs synthetic paired supervision dataset is: starting from real videos with diverse camera motion and pose annotations, treating each video clip as a 'geometrically consistent target', and then using a pretrained image-to-video restoration model with trajectory-adaptive keyframes to synthesize the corresponding 'distorted version'.69% similarUnverifiedThe generated videos and simulator labels together form a reusable multimodal memory bank, supporting a training-free multi-agent VLM system.67% similarUnverifiedCurrent mainstream VLA models typically concatenate multiple RGB images directly or input them to the policy network in a fixed order, implicitly assuming that the number of cameras, mounting positions, and intrinsics and extrinsics at training and test time are completely identical.66% similarUnverifiedThe vast majority of mainstream VLMs (such as LLaVA and various open-source multimodal models) use CLIP or its variants as the visual encoder.66% similarUnverifiedSmolVLA is a lightweight vision-language-action (VLA) model released by HuggingFace, with parameter count kept within a range that allows it to run and be fine-tuned on consumer-grade GPUs.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1073787API
curl https://kongchang.com/api/v1/knowledge/claims/1073787MCP
get_claim(id=1073787)