How well do leading VLMs understand the physical world?
Introducing the Encord Egocentric VQA Benchmark, new research from Encord Data Labs comparing VLM performance with professional human annotators.
Built from 1,500 questions on unseen, real-world video, the benchmark reflects the types of scenes and captions we provide to our robotics customers.
Today, humans still lead. A single annotator outperformed Astra, the best model tested, by 8 percentage points. Hybrid human–VLM annotation also showed promise.
One of the most surprising findings was on consensus.
Majority voting across three human annotators improved accuracy by around 5 percentage points, while combining VLMs provided no improvement over the best individual model. In some cases, it actually degraded performance.
Explore the results and implications for Physical AI data development in our latest research article. Link in comments.
Introducing Universal Manipulation Exoskeleton (UME)
A low-cost exoskeleton with real-time haptic torque feedback for learning autonomous policies that perform highly force-mediated, tightly space-constrained, visually occluded, whole-body, and long-horizon mobile manipulation tasks.
Using UME, the teleoperator can unsheathe a heavy metal sword completely blindfolded.
https://t.co/W3PHmYRm4q
🧵1/N
The Band of the Coldstream Guards played ‘We’ll Meet Again’ in London this afternoon, in tribute to Dame Vera Lynn. The Coldstream Guards also provided a ceremonial Guard of Honour for President Macron’s visit to the UK today.