JADE
A vision-language-action (VLA) model architecture in which a pretrained VLM dynamically reads noisy action latents within a diffusion denoising loop and, through gated cross-attention, guides an action expert, enabling joint denoising and action generation.
Timeline (last 90 days)
JADE (Joint Action DEnoiser) is a VLA architecture method that lets the VLM and the action expert jointly participate in the action refinement process.
JADE lets the VLM read the current noisy action latent variable at each denoising step and contextualize it together with multimodal inputs.
Through a gated cross-attention mechanism, JADE lets the VLM dynamically adjust the weight of the guidance signal according to the current noisy action state.
JADE's training objective includes supervising the VLM to recover clean action latents from noisy action latents.
JADE can generate actions with fewer sampling steps, thereby partially offsetting the additional computational cost brought by the VLM's involvement.
JADE introduces an adaptive step-size scheduling strategy that balances task performance and inference latency.
On the RoboDojo benchmark, JADE achieves an average success rate of 12.67%.
On the RoboDojo benchmark, JADE ranks highest among all evaluation methods based on π₀.₅.
On RoboTwin 2.0, JADE likewise brings performance improvements.
Across four real-world dual-arm manipulation tasks, JADE improves the average success rate by 20.8 percentage points compared with π₀.₅.
All Facts (10)
JADE (Joint Action DEnoiser) is a VLA architecture method that lets the VLM and the action expert jointly participate in the action refinement process.
50%UnverifiedJADE lets the VLM read the current noisy action latent variable at each denoising step and contextualize it together with multimodal inputs.
50%UnverifiedThrough a gated cross-attention mechanism, JADE lets the VLM dynamically adjust the weight of the guidance signal according to the current noisy action state.
50%UnverifiedJADE's training objective includes supervising the VLM to recover clean action latents from noisy action latents.
50%UnverifiedJADE can generate actions with fewer sampling steps, thereby partially offsetting the additional computational cost brought by the VLM's involvement.
50%UnverifiedJADE introduces an adaptive step-size scheduling strategy that balances task performance and inference latency.
50%UnverifiedOn the RoboDojo benchmark, JADE achieves an average success rate of 12.67%.
50%UnverifiedOn the RoboDojo benchmark, JADE ranks highest among all evaluation methods based on π₀.₅.
50%UnverifiedOn RoboTwin 2.0, JADE likewise brings performance improvements.
50%UnverifiedAcross four real-world dual-arm manipulation tasks, JADE improves the average success rate by 20.8 percentage points compared with π₀.₅.
50%