Nori Robotics Targets Developers With a $1,688 Bimanual Mobile Robot

Introduction

For many robotics teams, hardware availability is as important as algorithmic performance. Nori Robotics is positioning its Nori A3 as an inexpensive development platform rather than as a finished general-purpose household robot. In a Hacker News launch post, the company says the bimanual mobile robot costs $1,688, is assembled in San Francisco, and is expected to ship in fall 2026.

The central proposition is straightforward: make a capable-enough robot cheap enough that labs and independent developers can buy more than one. That could support larger demonstration datasets, longer experiments, and repeated tests across multiple machines—activities that are difficult when a research group has only one or two expensive robots.

Key specifications and design choices

• Manipulation and mobility: Nori lists 19 degrees of freedom, with two 7+1-DOF arms. Each arm is rated for a 1.5 kg payload. The robot moves on a differential wheeled base rather than legs.
• Perception: Four 720p RGB cameras can capture up to 30 frames per second. They are mounted around the grippers, head, and neck. A 2D lidar is also included; the website lists a 12-meter range and an 8–12 Hz scanning frequency.
• Interaction and battery: The platform includes a speaker and microphone system for spoken commands and full-duplex voice communication. Its 432 Wh battery is advertised with six to eight hours of battery life.
• Computing architecture: A Raspberry Pi 5 with 4 GB of RAM runs SLAM and safety functions onboard. More demanding ACT and vision-language-action models must run on a computer over a LAN connection or on a server over a WAN connection.
• Cost reduction: Nori says it chose high-ratio servos instead of more expensive quasi-direct-drive motors and used wheels instead of legs. The company also says the robot has more than 100 moving and structural parts and is designed to be manufactured and repaired more easily, with 3D files made available.

What the price does—and does not—mean

The $1,688 price is notable because it shifts the discussion from the capability of a single premium robot to the number of platforms a team can deploy. More units can make it easier to collect demonstrations, run parallel trials, and investigate how a policy behaves across hardware. For researchers working on imitation learning or robot control, that quantity can be as valuable as a higher-end actuator or sensor.

The trade-offs are equally important. A wheeled base should be easier and cheaper to build than a legged system, but it cannot provide the same access to stairs or uneven terrain. Offloading advanced models to a laptop or server reduces onboard hardware costs, while introducing dependence on network connectivity and external compute. The cameras, lidar, and arm payload also describe a development-oriented configuration, not a demonstrated guarantee of robust household autonomy.

Nori’s website presents home tasks such as fetching, dish loading, folding clothes, and pouring ingredients, as well as a skills marketplace and a laptop application for training and operation. These are product ambitions rather than evidence that the robot already performs each task reliably. Public information currently provides specifications and a manufacturing rationale, but not detailed benchmarks for manipulation success, uptime, latency, or delivery performance.

That leaves the software ecosystem as a decisive factor. If the Nori Lab application, teaching workflow, repair process, and shared skills work as intended, the platform could become useful precisely because it is repeatable and accessible. If not, a low purchase price alone will not solve the practical barriers to embodied-AI research.

Nori A3 therefore represents a clear engineering compromise: use wheels, high-ratio servos, and remote compute to bring a bimanual robot below $2,000. It does not show that general-purpose robotics has been solved, but it does illustrate a potentially important path for the field—more affordable hardware, deployed in greater numbers, generating the data needed for better models.

Source: Hacker News