Learning from an Imitable Teacher: Competition Solution for the MetaFine Challenge
Learning from an Imitable Teacher: Competition Solution for the MetaFine Challenge
This paper presents Team Friday's second-place solution (score 0.947) for the MetaFine Open Competition at CoRL 2026, which focuses on fine-grained manipulation tasks in simulation. Instead of fine-tuning a large vision-language-action model, the team trains one 19M-parameter policy per task with a separate instruction classifier. The approach utilizes demonstration data generated by a privileged scripted teacher that corrects student mistakes. Analysis reveals that a teacher that almost always succeeds can teach a memoryless student poorly due to pauses becoming stalls and fixed habits becoming shortcuts. By making the teacher imitable, the score on the hardest letter insertion task improved from 0.44 to 0.88 with 32% more data.
The authors present a competition solution for the MetaFine Challenge that achieves a second-place score of 0.947 by training task-specific 19M-parameter policies using data generated by an imitable scripted teacher, rather than fine-tuning a large VLA model.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S2
- Control
- ScriptedOpen source S3
Reported numbers
overall score
0.947
View original evidence
am Friday's solution, which got second place with an overall score of 0.947, more than twice the $\pi_0$ baseline. Instead of fine-tuning a large vision-language-action (VLA) model, we train one 19M-parameter policy per task with a separate instruction classifier and focus on the data, all generated in the simulator by a privileged scripted teacher, including its corrections of the student's mistakes. By analyzing fa
Open source S3letter insertion score
0.44–0.82
View original evidence
iled episodes, we find that a teacher that almost always succeeds can still teach a memoryless student badly: its pauses become stalls, its fixed habits become shortcuts, and it never shows the errors the student makes. Making the teacher \emph{imitable} raised the official score on letter insertion, the hardest task, from 0.44 to 0.82 with 32\% more data and the same architecture, and finally to 0.88. We summarize p
Open source S4letter insertion score
0.88
View original evidence
iled episodes, we find that a teacher that almost always succeeds can still teach a memoryless student badly: its pauses become stalls, its fixed habits become shortcuts, and it never shows the errors the student makes. Making the teacher \emph{imitable} raised the official score on letter insertion, the hardest task, from 0.44 to 0.82 with 32\% more data and the same architecture, and finally to 0.88. We summarize p
Open source S4baseline score
0.44–0.82
View original evidence
iled episodes, we find that a teacher that almost always succeeds can still teach a memoryless student badly: its pauses become stalls, its fixed habits become shortcuts, and it never shows the errors the student makes. Making the teacher \emph{imitable} raised the official score on letter insertion, the hardest task, from 0.44 to 0.82 with 32\% more data and the same architecture, and finally to 0.88. We summarize p
Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
arXiv:2611.00615v1 Announce Type: new Abstract: The MetaFine Open Competition (RoboFineMani workshop, CoRL 2026) evaluates fine-grained manipulation beyond binary success. Its five simulated tasks, from grasping a bottle by an instructed part to inserting letter-shaped pegs with 4 mm clearance, are scored on perception under lighting and camera changes, instruction understanding, and behavior. This report presents Te
Open source S2
am Friday's solution, which got second place with an overall score of 0.947, more than twice the $\pi_0$ baseline. Instead of fine-tuning a large vision-language-action (VLA) model, we train one 19M-parameter policy per task with a separate instruction classifier and focus on the data, all generated in the simulator by a privileged scripted teacher, including its corrections of the student's mistakes. By analyzing fa
Open source S3
iled episodes, we find that a teacher that almost always succeeds can still teach a memoryless student badly: its pauses become stalls, its fixed habits become shortcuts, and it never shows the errors the student makes. Making the teacher \emph{imitable} raised the official score on letter insertion, the hardest task, from 0.44 to 0.82 with 32\% more data and the same architecture, and finally to 0.88. We summarize p
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org