E-TTS: A New Embodied Test-Time Scaling Framework for Robotic Manipulation
Wen Ye, Peiyan Li, Tingyu Yuan, Yuan Xu, Xiangnan Wu, Chaoyang Zhao, Jing Liu, Nianfeng Liu, Yan Huang, Liang Wang
Read on arXiv →
Key claim
E-TTS improves robotic performance by integrating reasoning and history.
In plain English
Imagine you're building a robot that needs to perform tasks in a complex environment, like picking up objects or navigating through obstacles. These tasks often require the robot to make decisions based on a sequence of actions over time, which can be tricky because it needs to remember what happened earlier to make the best choice now. This is where the challenge lies: relying only on what the robot sees at the moment can lead to poor decisions, especially if it forgets important past information. This issue is known as the lack of historical context utilization.
Current methods often struggle with this because they don't effectively incorporate past experiences into their decision-making process. They might use a straightforward approach where the robot acts based solely on its current observations, which can lead to mistakes when the situation is complex or requires long-term planning. This is what's called open-loop test-time scaling, where the robot doesn't adapt its actions based on feedback from its environment.
The E-TTS framework addresses these challenges by combining reasoning and action scaling in a way that allows the robot to learn from its past actions. It uses a history buffer to keep track of previous states and decisions, which helps the robot evaluate its options more effectively. By introducing a feedback loop into the decision-making process, E-TTS enables the robot to refine its actions iteratively, improving its adaptability and efficiency in real-time.
Compared to previous methods, E-TTS not only enhances performance but does so without needing additional expert data or retraining. This means that for builders working on robotic systems, implementing E-TTS could lead to more capable robots that perform better in dynamic environments, making them more useful in practical applications.
The paper introduces a new framework that effectively combines reasoning and action scaling in embodied tasks, addressing previously unsolved challenges.
The experimental results are based on multiple benchmarks and environments, supporting the claims made about performance improvements.
Deep reliability assessment
The evidence supports that E-TTS can improve success rates when wrapped around several VLA models across multiple simulated and real manipulation settings. The stronger claim that this is a generally plug-and-play embodied test-time scaling solution is less fully supported without clearer compute/latency tradeoff curves, verifier generalization tests, and details on the action preference dataset used to train the verifier.
Reproducibility
A project page is provided at https://27yw.github.io/E-TTS-Web/, but the provided text does not explicitly mention released code, model weights, or the action preference dataset.
Key figure
Figure 1 contrasts prior action-only test-time scaling with E-TTS, which adds reasoning, history, feedback, and verifier-based candidate selection across multiple VLA models, environments, and robot embodiments.