Rhoda AI

Yesterday, our research team gathered with academic colleagues over drinks at #CVPR to explore the future of video models and robotics. One theme kept coming up: excitement around world models for robotics. The interest from the research community has been incredible — we're entering a new phase for Physical AI, driven by real-time video models. We are witnessing conversations at the frontier of this technology that are truly unprecedented. A sincere thank you to the incredible group of visionaries and researchers who joined us. The turnout was amazing! Huge shoutout to the Rhoda team for making it happen. If you're here, come say hi to Eric Chan, Alexander Bergman, Ivan Skorokhodov, Qianqian Wang, Gordon Wetzstein, Vincent Sitzmann, and Alan Baade.

Can a large foundation video model run as a real-time robot policy at the edge, on a single RTX 5090? • ✅ No quantization • ✅ No distillation • ✅ Full denoising (all the way from noise to clean video) We just proved it's possible. 👇🎬 How? Existing video models aren't optimized for real-time inference. Instead of fine-tuning off-the-shelf video models, we co-design inference-aware model architectures and model-aware inference optimizations from the ground up.

Teaching a robot a new task typically means stopping operations, collecting teleoperated demonstrations, and retraining. That process takes hours at a minimum. We wanted to know if we could collapse it to seconds — from a single human demo, on the fly, no retraining required. Early research preview: we can. The hardest version of this problem is what we went after. Human demos with no action labels. Different camera views between the demo and execution. Following intention and motion, not just copying keyframes or a final state. We stress-tested on item sorting and drawing. How it works: we train on paired human demo and robot execution data. Because our DVA, FutureVision, has long-context visual memory built in, we prepend the full human video into the model's context and predict robot actions closed-loop. The model watches a human do something once and understands what to do next. The future we're building toward is one where robots adapt to new tasks in seconds. A human shows the robot what to do. The robot does it. At Rhoda, we tackle real-world problems through fundamental research. 

Our co-founder and CPO, Andrew Wooten , was invited to speak at Columbia Business School with the CBS Tech Club. Thanks to Alina East for moderating and to the club for having him.

Here's something we've never seen done before. Real-world tasks are long and ambiguous. Solving them requires visual memory, state tracking, and understanding how a scene evolves over time. Most robot policies see only the last few frames. Ours doesn't. We put our DVA, FutureVision, to the perfect testbed: the shell game. 🐚 An object is hidden under one of three identical cups. The cups are shuffled. The robot must find it with no direct line of sight, tracking every swap from the very start. A single lapse in attention means failure. Simple for humans, highly non-trivial for machines. This task is intractable for robot policies with little or no visual memory. Our DVA solves it with a 96%+ success rate across 24 consecutive successful reveals. But what stands out even more is the generalization: the model handles completely new objects, new numbers of shuffles, and even a different number of cups than it ever saw during training. How? Our DVA implements robot policy as future video generation. Given the context, the model generates future video predicting not just the correct cup to pick up, but even the appearance of the hidden object. Training natively on long, continuous video is what gives DVA built-in long-context memory — no separate modules, no scaffolding bolted on. At Rhoda, we tackle real-world problems through fundamental research. This is what that looks like. Full story and technical deep-dive: https://lnkd.in/gbVFaJCX

Most robot models are trained on task specific data in controlled environments. They perform well in the lab and struggle the moment conditions change on a real factory floor. Rhoda's approach starts differently. Web scale video pretraining builds a general prior on how the physical world moves. That foundation is what lets our DVA model handle broken latches, rusted tables, and deformable objects it has never seen before, with only a few hours of task specific data. Our CEO, Jagdeep Singh, spoke with Brian Heater at A3 about why this paradigm works and where it's headed. Full article in the comments.

We are speed running industrial robotics. It took us just 19 days from the first day of data collection to filming a 2.5-hour continuous run of our model autonomously breaking down industrial containers — zero human intervention. The data efficiency of our DVA model is fundamentally changing how fast we bring robots out of the lab and into the factory. Breaking down these containers requires more than just raw force. Each arrives in a random orientation with diverse, deformable debris inside. Because of the oversized dimensions, the model must constantly navigate partial observability. Success depends on accurate spatial reasoning — knowing exactly when to re-orient the box or pull it closer to reach far corners. Achieving a 100% autonomous rate over a 2.5-hour continuous run means the model must handle a wide range of edge cases. Whether it's pulling a drifted box back into range or re-attempting a failed flip, the model self-corrects in real-time. Here's what that looks like: → The trash is out of reach. The robot repositions the box before attempting another grab. → The door won't fall open. The robot recognizes a latch wasn't fully released and goes back to fix it. → The first flip fails. The robot doesn't hesitate — it goes for a second attempt. → The box has drifted too far to reach the latch. The robot pulls it back into range. At Rhoda, we solve real-world problems with fundamental research. Full story + technical deep-dive in the comments.

Most robot demos are “golden runs”: a perfect take selected from many attempts. But real-world deployment is about Continuous Operation. Watch our DVA model tackle a real-world decanting task for 1.5 hours straight: Uncut, Zero human intervention. -- Our industry partner thought this was “impossible to automate”. Why? 💥 Chaos is normal: straps rip, bearings jam, boxes topple 🏋️ Heavy lifting: 10kg boxes 🎯 Precision: needs to find and pull small tabs 🌊 Deformables: needs to handle plastic bags that never look the same twice -- Trained on just 11 hours of robot data, our model is surprisingly robust, thanks to web-scale pre-training. It doesn't just avoid errors; it handles them. If the lid tears off, it finds a new way to grip. If a bearing is stuck, it shakes the bag loose. -- https://lnkd.in/gGZN4pMR

Our CEO Jagdeep Singh talks about taking on the largest unsolved problem in technology.

Our model continuously and autonomously empties then breaks down 50-pound, heavy duty boxes common in large-scale manufacturing.