Config demonstrates long-horizon memory in CFG-1 robot foundation model
Evidence notes
- Config demonstrated long-horizon memory capabilities in CFG-1 across piano-playing, shell-game tracking and object-rearrangement tasks.
- The company reported 70 percent success on 8-note piano trials, 87 percent success in shell-game trials and 70 percent success on desk-rearrangement trials.
- Config framed the work as evidence that robot memory depends on collecting the right long-horizon training data.
Company context
Config builds robotics data infrastructure and foundation-model technology for bimanual manipulation. Its platform covers the workflow from task and gripper definition through human and teleoperation data acquisition, data validation, action estimation, model training and evaluation.