Many MindedRhoda AI trains robots on video to skip task-by-task programming
Rhoda AI, based in Palo Alto, California, emerged earlier in 2026 from 18 months in stealth with $450 million in Series A funding. Co-founder and CEO Jagdeep Singh says the company is working with customers in automotive, manufacturing, logistics and e-commerce to deploy its direct video action (DVA) model.
DVA takes a history of video frames, action tokens and language tokens, then predicts roughly the next half second of video. A second component, an action converter, turns that predicted scene into robot actions. The robot carries them out, observes again, and the loop repeats. Rhoda has also released a mobile version that combines manipulation and movement in a single policy.
The model is pretrained on video spanning hundreds of years of footage, not only footage of people. Rhoda says the broadest video sets beat curated human egocentric manipulation videos. After pretraining, a new task needs about 10 to 20 hours of post-training. The specific robot body matters only at that stage, so the same base model can move across embodiments. Rhoda always owns the AI layer, and it can run on its own hardware or a third party's.
Why it matters: the expensive part of automating physical work is usually bespoke programming for each task and machine. If a general model pretrained on abundant video can be adapted in under a day of training, the cost of putting robots into manufacturing and logistics, where Rhoda cites labor shortages, could fall. Video is also a far more plentiful teaching resource than robot-collected data.
Caveats: the performance and generalization claims come from Rhoda and its CEO, and no independent benchmarks are cited. Singh's statement that the number of autonomous manipulation robots solving real-world problems is effectively zero, and his prediction that robotics' ChatGPT moment will come sooner than expected, are his own assertions. Watch for named customer deployments and third-party testing.
Source: The Robot Report