From Needy Plant Pets to Algae Robots, AI Is Getting Physical

AI is stepping out of the phone and into the objects, robots, and materials around us. A tiny E Ink display can turn a houseplant into a needy virtual pet, while new robotic systems can learn physical skills, move with algae, or run specialized science models on local hardware.

That shift connects several projects that look different at first. They all give technology a stronger link to the physical world, whether through sensors, living cells, simulated environments, or focused datasets. The result is AI and robotics that respond to what objects do, what environments require, and what people want to accomplish.

Everyday objects become expressive companions

Daehwa Kim, a fifth-year PhD student in Carnegie Mellon’s Future Interfaces Group, created a system called OATs, short for Object Agent Twins. These micro-agent AI avatars appear on tiny E Ink displays attached to household objects such as plants, skateboards, coffee machines, and dumbbells.

Forget to water a plant, and the device does not show yellowing leaves. Instead, a cute plant-shaped pet on the screen looks sad and dejected. Pour water into the soil, and the pet becomes excited and happy. Kim says, “We bring this joy of caring from virtual to physical objects.”

Kim presented OATs at the ACM Symposium on User Interface Software and Technology in Detroit in early November, where the project earned two Honorable Mention awards. The forum explores human-computer interfaces, but OATs move that interaction away from a phone screen and onto the things people already use.

Each OAT combines an E Ink display with internal and external sensors that can track moisture, temperature, humidity, light, occupancy, and motion. The main module costs around $20, and users set it up through a web portal where they name the OAT, choose sensors, and describe the object. The web app then creates “Positive” and “Negative” behaviors based on that information.

A plant that prefers dry conditions might treat consistent air temperature readings and low humidity as positive signals, though users can adjust the behaviors for a specific plant type. A skateboard avatar can look sad and lonely after several days without detecting motion, then show fun, happy faces when someone rides the board. Kim says, “These kinds of inanimate objects can be more expressive.”

The sprites come from Google Gemini’s Nano Banana image-generation service, which creates a range of emotions for different sensor states. People have also used OATs to answer three questions: “Am I being used?”, whether an object should warn people nearby, and whether it can help track personal behavior. A coffee machine can get excited when it grinds beans, an OAT on a 3D printer can show that the device is operating, and one attached to a dumbbell can track how often someone lifts weights.

The main limitation is battery life. OATs last around three to four days, or only a day when paired with the sound level sensor. They charge through a USB-C port connected to a lithium-ion polymer battery, making the system a proof of concept rather than a finished household product.

Robots learn, grow, and react

The OAT demonstration became popular enough that Kim turned it into a competition over the most interesting object pairings. It also raised a harder question: does giving a machine human-like emotion make people feel more connected to it? The same issue already appears in AI, where companies describe chatbot behavior with human terms such as “hallucinate.”

Another project shows how AI can train robots inside simulated spaces. Will Knight taught a virtual version of Microduck, a robot from Pollen Robotics and Hugging Face, to walk around a room and identify bengal cats using OpenAI’s GPT-6 Sol. The virtual robot learned the task in just a few hours, and Knight plans to have a physical Microduck dispense a treat to his bengal cat, Leona, by bowing its head toward a 3D-printed, head-mounted storage chamber.

The setup used Hermes from Nous Research, an Nvidia DGX Spark, and Nvidia Isaac Sim. GPT-6 Sol learned to move the virtual robot without toppling over, though an attempt to teach it a moonwalk used a GIF of Michael Jackson sliding across a stage. Knight says OpenAI’s models can understand and manipulate 3D representations of the physical world, then turn them into code, agentic actions, and robot training.

Ken Goldberg’s lab at the University of California, Berkeley, used Nvidia’s Newton tool after OpenAI released Astra this past September. When a lab member asked Astra to create a simulation, Goldberg says it produced a full 3D scene within an hour, including the camera position, lighting, background, and materials. Astra can also automate much of the work involved in training a robot in simulation.

Jim Fan, a researcher at Nvidia, says this resembles the way software agents use different programs. Large language models can work with visual language action models that control physical movements, just as software agents access different tools on a computer. Fan, Goldberg, and others open sourced Aspire, an agentic framework that lets robots discover and use new skills. Examples include GPT-6 Astra using robot arms to solve a Rubik’s cube and controlling a robot arm to paint the Golden Gate Bridge.

Reliable robots still require significant engineering expertise. The vision system on Knight’s robot duck is not ready for a real product, but easier robot training could affect manufacturing and e-commerce. Goldberg says industrial robotics and automation firms have shown interest in the approach.

Living materials and focused AI models

At Delft University of Technology in the Netherlands, Marie-Eve Aubin-Tam led research into soft robotic arms powered by algae. The tiny arms measure about three-quarters of an inch long and are about as thick as a credit card. Two hydrogel layers hold algal cells, and one layer responds to pH changes caused by the algae’s metabolism.

Light makes the algae photosynthesize, absorb carbon dioxide, and fix nitrogen. That makes the local pH more basic, causing the responsive hydrogel to swell. In darkness, the algae release carbon dioxide through respiration, the pH returns toward the acidic side, and the hydrogel shrinks. The different volumes of the two layers bend the arm.

The chemical shift starts about 20 seconds after light is applied and about 10 seconds after darkness begins, though the hydrogel takes several hours to bend significantly. At pH 3 to 7, the responsive layer measured approximately 37 percent of its original weight; at pH 11, it reached four times its original weight.

The team also built a four-fingered gripper that closed under light and lifted a hollow metal sphere, releasing it when the light switched off. A hook-shaped gripper picked up and placed down a rubber ring. The objects weighed far less than a small paperclip, yet the devices handled them with dexterity.

Aubin-Tam says, “We wanted to explore what becomes possible when living cells are integrated into functional materials.” The algae require no external energy source beyond light and no complex mechanical parts, which could support long-term operation with little outside input. Possible uses include automated agricultural equipment, fluid control in lab-on-a-chip devices, and underwater robotics.

On the data side, Tether’s Genesis III takes a different route toward focused machine intelligence. It is an open synthetic STEM dataset designed to train small models for scientific technology, engineering, and mathematics. The combined corpus contains 191.43 billion tokens across 159.6 million documents, building on Genesis I, released October 24, 2025, and Genesis II, released December 22, 2025.

The QVAC Genesis III paper was accepted and presented at the 2026 Conference on Language Modelling. Testing found that models trained on Genesis III performed better on science and STEM tests than models trained on comparable open-source datasets. A 1.7-billion-parameter model trained on one part of the dataset delivered valid answers in 99.45 percent of benchmark responses.

The larger idea is simple: not every useful AI needs a massive cloud model. Genesis III targets laptops, smartphones, and company servers, allowing specialized models to work without an internet connection. Alongside expressive objects, simulated robots, and algae-powered machines, it points toward technology that is more local, more responsive, and more connected to the physical world.

Stay connected via Google News

Follow us for the latest AI News