Task-Specialized Language-Conditioned Models for Fine-Grained Robotic Manipulation
Task-Specialized Language-Conditioned Models for Fine-Grained Robotic Manipulation
This paper presents a submission to the MetaFine Open Competition at CoRL 2026, featuring five independently trained models that utilize frozen DINOv2 and CLIP encoders combined with trainable language-conditioned feature modulation. The approach improves performance through contact-aware grasp filtering, instruction-matched data, and expert correction of insertion failures, achieving an official overall score of 0.914 and ranking third among 15 teams.
The authors report an official overall score of 0.914 and a third-place ranking among 15 teams in the MetaFine Open Competition at CoRL 2026, achieved by five independently trained models using frozen DINOv2 and CLIP encoders with task-specific action heads.
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.
Reported numbers
overall score
0.914
View original evidence
improve demonstrations through contact-aware grasp filtering, instruction-matched data, and expert correction of insertion failures. At inference, task-dependent strategies include photometric test-time augmentation, stochastic action averaging, and uncertainty-triggered replanning. Our submission achieves an official overall score of 0.914 and ranks third among 15 teams. We analyze the models and their task-specific
Open source S3
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
improve demonstrations through contact-aware grasp filtering, instruction-matched data, and expert correction of insertion failures. At inference, task-dependent strategies include photometric test-time augmentation, stochastic action averaging, and uncertainty-triggered replanning. Our submission achieves an official overall score of 0.914 and ranks third among 15 teams. We analyze the models and their task-specific
Open source S3
arXiv:2611.00955v1 Announce Type: new Abstract: Fine-grained robotic manipulation requires instruction grounding, robust perception, and precise contact control. We describe our submission to the MetaFine Open Competition at CoRL 2026. Five independently trained models use a common architecture with frozen DINOv2 and CLIP encoders, trainable language-conditioned feature modulation, and task-specific action heads. We
Open source S2
Task-Specialized Language-Conditioned Models for Fine-Grained Robotic Manipulation
Open source S1
behavior.
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org