GO-1 (Genie Operator-1, Zhiyuan Qiyuan) is Zhiyuan Robotics’ first general embodied base model — a single brain intended to power many different robot bodies and form factors. The architecture is called Vision-Language-Latent-Action (ViLLA), and it combines a VLM (vision-language model) with a MoE (mixture of experts) whose specialists handle planning and action.
The division of labor is the design’s substance. The VLM draws on massive internet-scale image-text data to give the model general scene perception and language understanding — it can identify what is in an image, fuse that with text input, and comprehend complex scenes end to end. The Latent Planner inside the MoE learns general action understanding from large volumes of cross-embodiment data and human-operation videos — absorbing how humans physically solve tasks. The Action Expert is trained on a million real-robot data trajectories and delivers fine-grained, efficient action execution on hardware.
GO-1’s Main Capabilities
- Learning from human video: by analyzing vast amounts of human-operation video, the model absorbs real-world action knowledge and adapts to new tasks quickly — human demonstrations become training data without robot-specific capture.
- Few-shot rapid generalization: with minimal data or zero samples, GO-1 generalizes to new scenes and tasks, lowering the application threshold that has kept embodied AI trapped in per-task engineering.
- One brain, many bodies: GO-1 deploys flexibly across many different robot types and body morphologies — cross-embodiment transfer is built into the architecture rather than bolted on per deployment.
- Continuous evolution: in production use, the model keeps learning and optimizing through a data-flywheel system that channels real-execution problem data back into training — it gets smarter with use, the dream metric of every robotics team since the field began.
- Efficient action execution: the million-trajectory-trained Action Expert gives the model refined, high-efficiency physical execution rather than approximate, jittery motion.
How the Computation Works
The three-stage pipeline reads like a translation chain from perception to physics. The VLM mines internet-scale image-text data for broad world knowledge: accurate recognition of image content, efficient fusion with text, and comprehensive understanding of complex scenes. The Latent Planner converts that understanding into action intent — having studied cross-embodiment and human video at scale, it holds general-purpose action-planning logic that transfers across arms, hands, and body types. The Action Expert takes the latent plan and produces the actual motor program, its million-real-robot-data training ensuring the output is executable precisely rather than approximately. The layered design also explains GO-1’s transferability claims: perception and planning are embodied-agnostic by construction, while embodiment-specific skill lives in the expert layer where it can be retrained per body.
The ViLLA architecture’s real innovation is the middle layer, and it is worth pausing on why. Most robot foundation models treat video as a data problem — caption it, tokenize it, mix it into training. The Latent Planner instead treats human video as a planning corpus: it learns not just what actions look like but the intent structure connecting them, expressed as latent (rather than literal) action tokens that sidestep the embodiment mismatch between a human hand and a robot gripper. This is what makes the cross-embodiment claim structurally credible rather than aspirational — because the planner’s output is embodiment-agnostic by construction, the same latent plan can route to a wheeled base, a bimanual arm, or a humanoid’s expert layer without redesign.
The data-flywheel claim — continuous evolution from real execution — deserves equal scrutiny and equal credit. Every deployed robot generates failure data, and failure data is precisely what base models cannot synthesize convincingly: a simulation can model a cup’s physics but not the ten thousand surprising ways real kitchens defy expectation. By channeling real-execution problems back into training, GO-1 converts its deployments into a compounding data asset, which is the same economic logic that made search engines and recommendation systems unassailable in their eras. The honest caveats remain real: continuous learning in production is still genuinely hard to validate — a model that “gets smarter” can also silently regress — and few-shot generalization claims always look their best in carefully curated demo videos rather than messy production floors. But the direction is unambiguous — the industry’s center of gravity is shifting from programming robots to training them, and GO-1 is among the first open projects where the training-data flywheel is a designed system component rather than an aspiration. Teams evaluating embodied foundation models should benchmark it against their own tasks and watch the open-data releases — the AgiBot World public corpus is itself a contribution that lowers everyone’s barrier to entry.
Where It Fits in the Race
GO-1 lands in a field moving fast toward exactly this shape — general base models for embodiment, with Google, Figure, and Physical Intelligence all publishing along related lines. Zhiyuan’s differentiators are its data and its openness strategy: the million-real-robot-trajectory corpus behind the Action Expert comes from the company’s own fleet-scale deployment and data-collection operation, and the accompanying AgiBot World open project has made a substantial share of that data public. The strategic logic mirrors what played out in language models: whoever assembles the largest high-quality embodied dataset trains the most capable base model, and the base model — not any single robot — becomes the platform others build on. For robotics teams, GO-1 is worth watching as a test of the one-brain-many-bodies thesis; for the industry, its real significance is the data flywheel going public — continuous evolution from real execution means every deployment now feeds the next generation. Project page: agibot-world.com/blog/go1. GitHub: github.com/OpenDriveLab/AgiBot-World.
