Spotter: Let the Embodied Model Lead, and the VLM Reflect for It
Spotter: Let the Embodied Model Lead, and the VLM Reflect for It
This research introduces Spotter, a framework that reverses the typical role of Vision-Language Models (VLMs) in embodied AI tasks. Instead of having the VLM plan every step, the embodied model leads execution while a parallel VLM monitors for errors and intervenes only when failures are detected. The authors report significant improvements on benchmarks like RoboCasa and RoboTwin 2.0 using Qwen and GPT as the VLM components, with minimal latency overhead compared to standalone embodied models.
The paper proposes Spotter, a framework where an embodied model leads execution while a parallel Vision-Language Model (VLM) monitors for errors and intervenes only when failures are detected; reported improvements on RoboCasa and RoboTwin 2.0 are attributed to this architecture rather than independent verification.
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
additional time per episode
13–16
View original evidence
takes only 13 to 16 s longer than with the embodied model alone and about 70% less time than with a VLM-led baseline using the same model. Our code is available at https://github.com/zqc3117/Spotter.
Open source S6
- code: Available Artifact link Open source S6
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
it, and returns control. We run Spotter with Qwen and with GPT as the VLM, and both improve the embodied models; with GPT, Spotter improves Cosmos Policy and $\pi_{0.5}$ by 5.6 and 7.5 percentage points on RoboCasa, and raises $\pi_{0.5}$ from 47.2% to 57.0% on the Hard setting of RoboTwin 2.0 and from 53% to 83% on a real robot. Because the VLM steps in only when an error is confirmed, a successful episode with Qwen
Open source S5
takes only 13 to 16 s longer than with the embodied model alone and about 70% less time than with a VLM-led baseline using the same model. Our code is available at https://github.com/zqc3117/Spotter.
Open source S6
Source：arXiv Robotics — research abstracts · arxiv.org