Dream Labs Launches Dream-1: A General-Purpose Robot Model for Dual-Arm Grippers

The 'egg from chicken' dilemma in the robotics industry is becoming a breakthrough for general-purpose robot model startups.

On October 19 local time, Dream Labs, a research team formerly associated with NVIDIA's GR00T, ended its stealth mode, launched its first model Dream-1, and announced the completion of a 90,000,000-dollar seed round funding. The GR00T research team is a core R&D force within NVIDIA's GEAR (General Embodied Intelligence Agent Research) laboratory.

Dream-1 provides services via public API, with an official pricing starting at $3 per hour. It allows control of previously unseen dual-arm parallel gripper robots to perform tasks that have not been demonstrated, without the need for specialized training data collection for new robots.

Dream Labs is attempting to break a long-standing industry dilemma. In the past, robots had to be useful before people would buy them, and to be useful required a lot of data, which could only be generated after sufficient usage by enough people.

This time, Dream Labs aims to change the way robotic intelligence is acquired, making robots useful first through the use of general models, and then accumulating data through actual usage to drive continuous model evolution. However, the path from laboratory demonstrations to stable operation in complex environments still faces technical and commercialization challenges.

From 'Having Data First' to 'Being Able to Work First'

Although the robotics industry is often compared to autonomous driving, the latter can continuously collect road data through daily driving, while many robots lack such an advantage. Differences in the structure, grippers, cameras, and working environments of various robotic arms mean that data accumulated by one robot may not be directly applicable to another.

This data challenge is precisely the problem Dream Labs aims to solve with a general model. Their efforts are mainly based on two industry changes: low-cost robot hardware is becoming increasingly widespread, and foundational robot models are beginning to show potential for cross-task and cross-body generalization.

If a new robot can complete some practical tasks after being connected to the model, it would have the opportunity to enter real-world scenarios earlier, creating a flywheel effect of hardware proliferation, model usage, data accumulation, and model improvement.

The team published a paper titled “World Action Models are Zero-shot Policies” in February 2026, proposing WAM (World Action Model) and releasing the DreamZero model.

Unlike VLA (Visual-Language-Action models), which mainly generate actions based on visual and language instructions, WAM combines future video state prediction with robot action generation, aiming to learn about objects, actions, and how the physical environment changes from video.

The paper states that DreamZero can retain generalization ability for new tasks with about 30 minutes of new robot operation data adaptation. However, this does not mean that no new body data is needed at all; rather, a small amount of adaptation data may replace large-scale exclusive training.

Dream-1 further productized this approach. It outputs the end-effector pose, i.e., where the gripper moves and how it opens and closes, rather than directly outputting all joint instructions for a specific robot. This design aims to establish a common interface across hardware, simultaneously providing visual information through the head and side wrist cameras, and attempting to break through some existing solutions' limitations on specific grippers and wrist perspectives.

Evaluation scores do not equal universal usability

Dream Labs' published evaluations show that Dream-1 achieves an average task score of 59% per task, with a single task attempt cost nearly 1/16 of Gemini 4 Argon.

This result provides preliminary support for low-cost robot models, but benchmark scores do not equal task success rates in real-world scenarios such as factories or homes, nor can they directly prove that the model is suitable for long-term operation without supervision.

Additionally, Dream-1 currently mainly targets parallel dual-arm gripper operations and does not cover mobile, walking, dexterous hand operations, or full-body control of general-purpose robot systems. The model does not directly plan complete arm paths, so it may fail in cases involving obstacles, joint limits, or significant differences in kinematics.

To address this, the company offers fine-tuning as a service (FaaS). When zero-shot capabilities are insufficient, users can submit a small amount of demonstration data to obtain a customized model for a specific robot or task.

Within the industry, competition around foundational robot models has already begun. NVIDIA's GR00T focuses on humanoid robots and their development ecosystem; Physical Intelligence's π series aims for general operational capabilities; and Google DeepMind's Gemini Robotics explores the integration of vision, language, and robotic action.

Dream Labs' difference lies in its commercial delivery approach, encapsulating robot control capabilities into an hourly billed API, and using FaaS to address the adaptation needs of new hardware and new tasks.

As a startup, Dream Labs bets on the idea that robots will not suddenly no longer need data, but rather that general models can start adapting from a small amount of data. If this capability continues to improve, robot manufacturers can allocate more resources to hardware and scenario development, without having to rebuild an entire model training system for each device.

But moving from low-cost APIs to scaled applications requires overcoming several hurdles: reliability, security, real-time control, and cost-effectiveness. Whether a robot's 'brain' can become an infrastructure like a cloud computing interface, callable by users, will depend on its ability to perform tasks consistently, stably, and at a low cost over the long term in real-world environments.