ADAPT architecture breakdown: diffusion Transformer plus residual RL in four stages
breadli428 · x · 2026-11-03
The ADAPT project page details a four-module pipeline: (A) dataset construction via retargeting and tracking to get text-labeled state-action trajectories; (B) a diffusion skill prior pretrained with CLP text encoding and a Diffusion Transformer Decoder, denoising in closed loop while predicting actions and next states; (C) residual RL correction on top of the frozen diffusion model for robust prompt switching (stand/punch/kick); (D) Noise Steering RL to guide the frozen prior for style-conditioned task execution. By Yan Wu et al. with Marco Hutter's group at ETH Zurich, accepted to CoRL 2026, code coming soon.
More from Embodied
- ADAPT: ETH Zurich's 50Hz end-to-end text-driven humanoid control with diffusion priors — breadli428 · 2026-11-03
- Xiaomi open-sources Robotics-1, a robot foundation model trained on 100K hours of trajectories — CyberRobooo · 2026-11-03
- Factory photos show desktop robot modules in volume assembly, project hits production stage — victormustar · 2026-11-03
- Machenta demos instant robot learning from passersby at London King's Cross — Ed__Johns · 2026-11-03
- PRIMA animal 3D mesh recovery accepted to WACV2027 with 9K pseudo-3D dataset — TrackingActions · 2026-11-03
- Engineer Ports NVIDIA's Open-Source SONIC Model to AGIBOT X2 in 6 GPU-Hours — Scobleizer · 2026-11-03