💻 computer science

Nori Bot: A Sub-$1,000 Floor-to-Counter Mobile Manipulator

The paper introduces Nori Bot, a $947 open-source 17-DoF mobile manipulator that overcomes the limitations of existing low-cost platforms by featuring a 600mm vertical lift, autonomous proactive control via an OpenClaw runtime, and a software-based safety stack to prevent servo burnout.

Original authors: Antonio Li, Sungjoon Park, Wen Ni Chew

Published 2026-05-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Antonio Li, Sungjoon Park, Wen Ni Chew

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a robot that can do chores around your house, but instead of costing as much as a luxury car (like many current robots), it costs less than a high-end gaming laptop. That is Nori Bot.

The researchers at Columbia University built this robot to solve three specific problems that have kept cheap robots from being truly useful: they can't reach high shelves, they only work when you yell at them, and their cheap motors often burn out if they get stuck.

Here is how Nori Bot fixes those problems, explained simply:

1. The "Elevator" Arm (Reaching High and Low)

Most cheap robots have arms stuck at one height. They can pick up a toy from the floor, but they can't reach a coffee mug on a counter.

The Fix: Nori Bot has a special "elevator" built into its body. The entire arm assembly sits on a rail that can slide up and down 24 inches (600 mm).
The Analogy: Think of it like a scissor lift at a construction site, but tiny and attached to a robot's arms. It can lower itself to pick up trash from the floor and then rise up to put a book back on a high shelf.
The Trick: Usually, adding an elevator requires a second, expensive computer brain. Nori Bot uses the same cheap motor and controller for the elevator that it uses for the right arm. It's like using one remote control to drive both your car and its sunroof, saving money and complexity.

2. The "Smart Alarm Clock" (Proactive vs. Reactive)

Current cheap robots are like walkie-talkies: they only talk when you press the button. If you want them to make coffee at 8:00 AM, you have to be there to press the button.

The Fix: Nori Bot connects to a "proactive agent" (a smart software system). This system acts like a smart alarm clock that can trigger the robot on its own.
The Analogy: Instead of you having to tell the robot to tidy up, you can set a "cron job" (a scheduled task). The robot wakes up at 8:00 AM, sees the coffee maker, and starts making coffee without you saying a word. It can also react to calendar events, like cleaning a table the moment a meeting ends.

3. The "Soft Fingers" and "Safety Net" (Preventing Breakage)

The motors Nori Bot uses are very cheap plastic gears. If the robot tries to grab something and gets stuck, the motor keeps spinning, the gears strip (break), and the robot dies. Also, cheap robots usually don't know how hard they are squeezing.

The Safety Net: The researchers wrote a "safety net" software. If the motor senses it is stuck (like a car hitting a wall), it immediately stops pushing hard and gently backs off. They also added "brakes" in the motor's memory so it can't be told to move to a spot that would break it.
The Soft Fingers: To know how hard it's squeezing, they replaced the robot's hard plastic fingers with soft, squishy rubber ones (TPU).
The Analogy: Imagine trying to feel a grape with a steel hammer; you can't tell how hard you're pressing without crushing it. But if you use a soft rubber glove, you can feel the grape squish. Nori Bot uses this squishiness to "feel" how hard it is gripping. If the rubber stops squishing and the motor starts spinning fast, the robot knows it has dropped the object or is slipping, and it stops.

What Can It Actually Do?

The team tested Nori Bot with five different tasks that require moving up and down:

Re-shelving books (reaching a mid-height desk).
Picking up trash (reaching the floor).
Sorting laundry (reaching a basket on a cart).
Folding clothes (reaching a table).
Making coffee (an autonomous task triggered by a schedule).

They also trained the robot using "imitation learning." This is like showing the robot a video of a human doing a task (picking up a small motor) 30 times. The robot then learned to do it itself.

The Bottom Line

Nori Bot proves you don't need a $25,000 budget to build a capable robot. By using a clever "elevator" system, a "smart alarm clock" software, and a "safety net" for its cheap parts, they built a 17-jointed robot for $947.

The ultimate goal isn't just to have one cool robot, but to build hundreds of these cheap units so they can all share their learning data, making future robots much smarter and more capable.

Technical Summary: Nori Bot

Problem Statement
Capable mobile manipulators remain prohibitively expensive for widespread research and consumer adoption, with commercial platforms like Mobile ALOHA ($32,000), Hello Robot Stretch (~$25,000), and PAL Robotics TIAGo (> $200,000) dominating the market. While open-source initiatives have reduced costs (e.g., XLeRobot at $660, AhaRobot at $1,000–$ 2,000), existing sub-$1,000 platforms suffer from three critical limitations:

Fixed Vertical Workspace: Arms are fixed at a single height, preventing tasks that require reaching from the floor to a kitchen counter.
Reactive-Only Control: Systems act only upon explicit human prompts, lacking the ability to execute scheduled, recurring household automation tasks.
Servo Fragility and Lack of Force Feedback: Low-cost platforms relying on Feetech STS3215 servos are prone to gear stripping under sustained stall conditions. Furthermore, they lack protection mechanisms and continuous grip-force feedback, which are essential for robust imitation learning.

Methodology
The authors present Nori Bot, a 17-DoF dual-arm mobile manipulator built on the XLeRobot mechanical base but enhanced with specific hardware and software innovations to address the identified gaps.

Hardware Architecture:
- Z-Axis Lift: A 600 mm linear-rail lift (HGR20 rails, RM1605 ball screw) is driven by a single Feetech STS3215 servo (Motor ID 11) on the same half-duplex bus as the right arm. This design avoids the need for a second controller board or a separate stepper motor protocol, allowing the lift to be treated as a standard joint in the kinematic chain.
- Cost Efficiency: The total parts cost is $947 (including a $287 lift subsystem), approximately 3% of the cost of comparable commercial platforms.
- Human-Robot Interface: A 7-inch touchscreen displays an animated face to convey task progress, errors, and confirmation states, serving as a critical HRI component rather than mere decoration.
Compute and Orchestration:
- Thin-Client Architecture: The on-robot brain is a Raspberry Pi 4 (1 GB RAM) running only motor I/O and a WebSocket bridge. Heavy computation (planning, vision) is offloaded to an external agent.
- Proactive Execution: The robot is exposed to the OpenClaw agent runtime as a "skill manifest." This enables tasks to be triggered by cron jobs (scheduled), hooks (reactive to events), or heartbeats, moving beyond simple prompt-response interaction.
Safety and Sensing Stack:
- Servo Protection: A three-layer software stack prevents Feetech servo burn-out:
  1. Calibration Clamping: Ensures commanded positions never exceed calibrated joint limits.
  2. Stall Detector: Monitors current vs. position change; if a stall is detected, torque is reduced to prevent gear stripping.
  3. Firmware Backstops: EEPROM registers are configured with hard current limits and over-current protection times.
- Sensorless Force Sensing: By replacing rigid gripper fingers with soft Thermoplastic Polyurethane (TPU) fingers, the system converts the motor's current register data into a continuous grip-force signal. The compliance of the TPU finger transforms a discrete "stall step" into a monotonic "squeeze ramp," allowing the imitation learning policy to observe force as a continuous variable.
Imitation Learning:
- The system utilizes ACT (Action Chunking with Transformers) via the LeRobot stack.
- Policies observe RGB camera feeds, joint states, and the normalized grip-force signal.
- The action space consists of 6 DoF target joint positions for the right arm and gripper; the Z-axis and base movements are scripted by the agent, keeping the policy checkpoint small.

Key Contributions

Single-Bus Z-Axis Lift: A 600 mm lift driven by the same servo protocol as the arms, enabling floor-to-counter reach without additional controller boards or complex protocol integration.
Thin-Client Proactive Orchestration: Integration of a low-cost manipulator with the OpenClaw runtime, enabling cron-scheduled and event-driven physical tasks via a lightweight on-robot client.
Robust Safety and Force Feedback: An onboard software stack that prevents servo burn-out and recovers continuous grip-force feedback from standard motor current registers using compliant TPU fingers.

Results

Task Demonstration: The robot successfully performed five end-to-end tasks across the full Z-axis envelope: book re-shelving, trash-to-bin pickup (floor level), laundry sorting (cart height), cloth folding (counter height), and an autonomous "make coffee" task triggered by a cron job.
ACT Policy Performance: A policy trained on 30 demonstrations successfully grasped a small DC motor from a desk. The system demonstrated that the Z-axis is necessary for tasks requiring variable heights.
Protection Efficacy: Over four weeks of operation, the protection stack logged 14 calibration clamping events, 23 gripper stalls, 6 joint stalls, and 1 firmware over-current trip. Zero servos were lost during this period, compared to two servos lost in the six weeks prior to implementing these protections.
Force Signal: The TPU finger successfully generated a continuous force signal (normalized [0, 1]) that correlated with squeeze depth, distinguishing between successful grasps and slips.

Significance and Scope
The paper positions Nori Bot as a per-unit foundation for a broader goal: enabling the widespread deployment of identical, low-cost robots to pool teleoperation data for training generalist policies (referencing works like $\pi_0$ and Open X-Embodiment). The authors explicitly state that this paper does not implement cross-unit data pooling or the resulting policy improvements; rather, it delivers the hardware and software infrastructure required to make such pooling feasible. The significance lies in lowering the unit cost and setup friction to a point where identical units can exist "in the wild," addressing the specific limitations of fixed height, reactive control, and hardware fragility that have hindered the sub-$1,000 tier.