Robotics / Deep Learning
A lightweight multimodal control policy that applies shared recursive reasoning to language-conditioned robotic manipulation.
View repository10
LIBERO tasks
16×
recursive passes
7-DoF
robot actions
This project explores whether a compact network can replace brute-force scale in vision-language-action modelling. A Tiny Recursive Model repeatedly refines the same hidden state, sharing parameters across passes while conditioning on camera observations, proprioception, and natural-language instructions.
Modern VLA policies are capable but extremely large and data-hungry. The goal was to preserve multimodal reasoning and precise continuous control while building a policy that can learn from the comparatively small demonstration sets available in robotics.
The resulting TRMPolicy combines pretrained visual and text encoders with a parameter-shared recursive module and a seven-dimensional action head. It is trained and evaluated on ten LIBERO-Spatial pick-and-place tasks using roughly fifty demonstrations per task.
Technical approach
Encode workspace images with ResNet18 and language instructions with a frozen CLIP text encoder.
Fuse perception, language, and proprioceptive state into a compact 256-dimensional representation.
Apply the same residual reasoning block recursively before predicting normalized seven-degree-of-freedom actions.