Dream Labs宣布推出机器人控制系统Dream-1及其公共API，该系统声称能够在零示例条件下控制未见过的双臂平行夹持机器人执行陌生任务，无需微调或机器学习专业知识1。该公司已完成9000万美元种子轮融资1，基于世界动作模型技术实现了早期零样本具身泛化的突破1。
Dream-1的兼容范围广泛，可控制任何配备双臂平行夹持的机器人，包括18年前发布的PR2型号1。该系统提供公共API接口，Dream-1版本的使用成本为每小时3美元，而Dream-1-flash版本为每小时2美元，每次调用返回1.6秒的动作序列1。为降低使用门槛，Dream Labs推出了Fine-tuning-as-a-Service（FaaS）功能，允许用户仅需5至98分钟的数据即可将Dream-1适配到特定机器人1。
根据RoboCurve的独立评估，Dream-1相比Gemini 4 Argon的成本低16倍，执行速度比GPT-6 Astra快1.7倍1。支撑这一系统的DreamZero论文已于2026年2月发表，阐述了World Action Models这一新型技术范式1。双臂机器人设置成本约为2000美元1。
Dream Labs has unveiled Dream-1, a robot control system equipped with a public API that enables zero-shot operation of unseen dual-arm parallel gripper robots performing novel tasks without fine-tuning or machine learning expertise1. The company simultaneously announced the completion of a $90 million seed round1.
The system leverages World Action Models technology to achieve early-stage zero-shot embodied generalization1. Dream-1 can control any robot featuring dual-arm parallel grippers, including legacy systems like the PR2 released 18 years ago1. The API is priced at $3 per hour for Dream-1 or $2 per hour for the Dream-1-flash variant, with each call returning 1.6 seconds of action sequences1. According to an independent assessment by RoboCurve, Dream-1 operates at 16 times lower cost than Gemini 4 Argon and 1.7 times faster than GPT-6 Astra1.
The platform includes a Fine-tuning-as-a-Service feature that allows users to adapt Dream-1 to specific robotic systems using datasets ranging from 5 to 98 minutes1. The underlying research was published in the DreamZero paper in February 2026, introducing the World Action Models paradigm1. The accessibility of dual-arm robot setups, with entry-level configurations costing approximately $2,000, has been bolstered by emerging affordable hardware like Hugging Face's $399 Microduck robot, which sold over 10,000 units within four days of its late-August launch1.
评论
还没有评论，欢迎留下第一条。