E-TTS: A New Embodied Test-Time Scaling Framework for Robotic Manipulation

Robotic manipulation models often struggle with long-horizon tasks because they rely on single-pass inference and lack mechanisms to correct errors or leverage historical context. The authors investigate how to apply test-time scaling (TTS) to the embodied domain to improve success rates without requiring additional expert data or model retraining.

The authors introduce E-TTS, a plug-and-play framework that treats reasoning and action generation as a unified, iterative process. Instead of scaling only the action space, E-TTS performs joint sampling of reasoning-action pairs. These candidates are evaluated by a dual-verifier system: a reasoning verifier that assesses semantic and spatial coherence, and an action verifier that evaluates physical feasibility.

Crucially, E-TTS incorporates a history buffer to provide temporal context, allowing the verifiers to make decisions based on the sequence of past actions and observations. If the top-scoring candidate fails to meet a confidence threshold, the system triggers a feedback-guided refinement loop, where the model receives textual feedback to generate improved candidates in subsequent sampling rounds.

E-TTS demonstrates significant performance gains across diverse settings, including four benchmarks, six environments, and three different robot embodiments. By integrating E-TTS with existing vision-language-action (VLA) models, the authors achieved up to a 33.14% increase in success rates in simulation and 26.62% in real-world scenarios. The ablation studies confirm that both reasoning scaling and action scaling are essential for optimal performance, and that the history-aware feedback loop is critical for navigating complex, sequential tasks.

This work provides a scalable, model-agnostic way to enhance the reliability of robotic agents. By shifting the burden of performance from expensive training-time data collection to inference-time computation, E-TTS offers a practical path toward deploying more robust robots in complex, real-world environments where task success is prioritized over millisecond-level latency.