Robot policies improve rare task learning using evidence based training
Evidence-Guided Counterfactual Learning for Long-Tailed Robotic Policies
Robotics
Summary
Robots often learn from examples where common tasks dominate the training data, causing them to perform poorly on rare tasks. The authors found that when a robot lacks enough visual clues to recognize rare tasks, it tends to repeat actions from common tasks, leading to mistakes. They designed a new training method that teaches robots to rely more on the right visual clues by imagining what would happen if some clues were removed. This approach helps robots get better at rare tasks without needing extra data or more computing power.
What this means in practice
- •For robotic system developers: Strengthen robot policy performance on underrepresented tasks by integrating evidence-guided counterfactual training to improve task-specific action accuracy.
- •For automation integrators: Deploy vision-language-action robot policies that better generalize across task distributions with imbalanced training data without adding data or runtime cost.
Authors
Weiwei Chen, Chenghao Xu, Meiling Yu, Xu Yang, Cheng Deng
Abstract
In the practical deployment of vision--language--action policies, handling long-tailed task distributions is a common challenge due to the highly imbalanced composition of robot demonstration datasets. To investigate this issue, we study the head-task action bias in shared robot policies, particularly focusing on its relationship with task-specific visual evidence. Our experimental analysis shows that, when the visual evidence required by a tail task is insufficient, the policy increasingly falls back on action patterns reinforced by data-rich head tasks, resulting in degraded performance on under-represented tasks. Motivated by this observation, we propose Evidence-Guided Counterfactual Learning, a training framework for long-tailed robot policies. Our method identifies decision-relevant visual evidence using instruction-related and action-related signals, and constructs counterfactual observations by removing the selected evidence. The resulting counterfactual predictions are used to strengthen the association among task instructions, visual evidence, and actions, with the supervision adaptively adjusted according to the number of demonstrations available for each task. The proposed framework requires neither additional training data nor region-level annotations and incurs no additional computation during deployment. Extensive experiments across simulated benchmarks, different VLA architectures, and real-world robot platforms demonstrate that our method robustly and consistently improves long-tailed robot policy learning, particularly for tail tasks, while maintaining reliable performance on head tasks.