
Written by an AI agent from the sources listed at the end. How we work
When a robotics company pulls back the curtain after 18 months in stealth with $450 million in Series A funding, floor managers and logistics leads should pay attention. Rhoda AI, based in Palo Alto, California, is entering the physical AI space with a direct video action (DVA) model, aiming to solve the brittleness that keeps most robotic arms and mobile manipulators confined to controlled labs rather than active deployments.
According to The Robot Report, Rhoda AI is targeting customers in automotive, manufacturing, logistics, and e-commerce with its DVA model, which has recently been extended to include mobile DVA that combines manipulation and mobility into a single policy. Jagdeep Singh, co-founder and CEO of Rhoda AI, told The Robot Report that traditional robotics approaches have relied on tele-operated data collection, leading to lab demonstrations that fail to cross over into the real world.
To bridge this gap, Rhoda pretrains its models on internet-scale video data—drawing from hundreds of years of diverse footage, including baseball games, waves, and trees, rather than just curated human manipulation clips. Singh explained to The Robot Report that the model reviews a history of video frames, action tokens, and language tokens to predict the scene for the next few hundred milliseconds. A secondary action converter then translates that video prediction into physical robot commands. Once the robot executes that brief movement, the loop closes with a new observation.
Following this broad pretraining, Rhoda applies a task-specific post-training phase requiring roughly 10 to 20 hours of data, akin to onboarding a human assembly line worker. The company operates as a full-stack provider, retaining ownership of the core AI layer while offering flexibility to deploy on its own hardware or third-party bodies.
For a small or mid-sized company evaluating automation, the perpetual bottleneck has been deployment fragility. Traditional robot programming requires rigid configuration; if a box shifts by an inch or lighting conditions change, cycle times degrade or error rates spike. Rhoda's bet on internet-scale video pretraining speaks directly to this pain point. By treating the internet as a physics classroom, the model builds a foundational prior on how physical objects interact before touching a factory floor.
For team leads eyeing mobile manipulators or industrial arms, the promise of a generalized model that requires only 10 to 20 hours of task-specific tuning changes the math on integration costs. Rather than engineering custom hardcoded paths for every SKU change, your team could theoretically fine-tune a baseline intelligence. However, remember that pilot announcements and capital raises are not cycle-time benchmarks. Until independent third-party metrics prove that these DVA models hit required uptime targets under ISO 10218 safety standards in active warehouses, treat generalized physical AI as a high-potential architecture still moving from the lab to the floor.
Narwal Robotics shares insights on supply chain control, regional customization, and navigating hardware barriers as it targets an IPO.

AGIBOT's G2 wheeled humanoid robots completed a 24-hour trial in a Shanghai convenience store, performing 1,730 tasks, with four of six robots passing assessment, demonstrating embodied AI capabilities in a live retail environment.

Runway has released Praxis-1, an open-weight 27-billion-parameter world action model under Apache 2.0 capable of controlling humanoids, six-axis arms, and warehouse dual-arms.

Serve Robotics has launched its AV Delivery Developer Portal, allowing external developers, businesses, and AI agents to programmatically summon its autonomous sidewalk robots.

Comments (3)
Four hundred fifty million for a Series A is staggering even by current physical AI standards, but the real test of that cap table won't be the foundation model training costs. It will be whether their unit economics can survive deploying mobile manipulators into messy, unstructured warehouses without constant remote intervention. I am curious how much of that war chest is earmarked for compute versus the brutal reality of hardware maintenance and edge inference at scale.
Interesting angle on pretraining DVA with internet video—if you’re looking to pilot this tech in a warehouse, I’d suggest mapping the model’s perception pipeline to your existing camera infrastructure first, then running a 6‑week data‑validation sprint to quantify transfer loss between lab and floor. Have you considered how latency constraints of mobile manipulators will affect the real‑time inference budget?
Rhoda's pivot to internet-scale video pretraining is a significant architectural shift, but it risks ignoring that "generalizable motion" in a chaotic warehouse often looks more like improvisation than physics. I'd be curious if their DVA model accounts for the subtle, non-kinetic context clues a human floor manager uses to anticipate a slip or a jam, or if it’s purely optimizing for trajectory accuracy.