cocacola-lab / Recursia

# Recursia: Recursive Self-Improvement for Robot Manipulation
Project Page  |  Paper  |  Real-Robot Quick Start  |  Simulation
Recursia improves a shared executable Python stack from source-task feedback,
then reuses the frozen stack on new task compositions. Improvements persist as
programs rather than model-weight updates.
This repository currently provides experiments on a Franka FR3 with a Franka
Hand, including a Skill Document Only (SD-only) baseline. Simulation code will be added under `recursia/sim/`.
## Quick Start
Create the environment for the Franka FR3 workspace:
 [code] 
Follow the real-robot setup guide to configure your
station and install/start SAM3 and Contact-GraspNet. Once ready, run the first
source chunk:
 [code] 
See Run for the remaining chunks, SD-only baseline,
and target evaluation.
## Demonstrations
Selected successful Recursia target executions on Franka FR3. Videos play at
original speed and concatenate recorded action segments; unrecorded perception
and model-wait intervals are omitted. Click a preview to open the full MP4.
Animated previews below run at 4x speed.
 Task Composition Video 
 T1: Shared basket 
 Place both the mug and the carton of milk in the basket. 
 Watch T1 
 T2: Reassigned plates 
 Place the blue mug on the left plate and the white mug on the right plate. 
 Watch T2 
 T3: Bowl--mug dependency 
 First place the bowl on the plate, then place the mug in the relocated bowl. 
 Watch T3 
## Repository layout
 [code] 
Logs, raw experiment recordings, model checkpoints, local station configurations, papers, and
independent experiment protocol documents are excluded from this repository.
Runtime prompt code required by the agents is included.
## Contribution Guidelines
See CONTRIBUTING.md. Run the offline tests from `recursia/real/` when
changing the physical runner. Keep simulation dependencies and coordinates
separate from real-robot configuration.
## Contact
For questions about Recursia, please contact
wjhan@bjtu.edu.cn or
chenyangzhi@bjtu.edu.cn.
## License
Inherited ASPIRE copyright notices are retained. Project material is available
under the Apache License 2.0, subject to component-specific terms.
See NOTICE and the license inventory.
## Citation
This implementation builds on ASPIRE. Please acknowledge the upstream work
when using its inherited components:
 [code]