Industries / Industrial equipment and services / Robotics
Teaching robots
Teaching robots is telling a robot what to do: guiding its arm through points, writing a program in a robot language, rehearsing the job in a simulator, or training an AI model on recorded demonstrations of the job.
6
Milestones achieved, from 1961 to 2025.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1961
Taught by leading the arm
The first Unimate, which went to work at General Motors' Ternstedt plant in Trenton, New Jersey, in 1961, had no programming language. An operator moved the machine's head to each spot, a position detector recorded it, and the robot played the recorded sequence back.
Credit
Unimation
1979
A language for robot arms
VAL is a programming language for robot arms, so that a job could be written down as a program rather than only recorded point by point. B. Shimano described VAL as "a versatile robot programming and control system" at the IEEE COMPSAC conference in Chicago in 1979, and Unimation used it on its PUMA arm. Sources disagree on its start: Robotiq's history says it was developed at Stanford in 1973 for Victor Scheinman's Vicarm robots and reached industry when Unimation bought Vicarm in 1977, while IEEE Spectrum says VAL was part of Unimation's robots from 1973. Robotiq says its descendant, VAL 3, is still in use.
LessonEach maker went on to use its own language (KUKA uses KRL, FANUC uses KAREL and TP, ABB uses RAPID, as a 2026 Chalmers University thesis lists them), so a program written for one brand of robot does not run on another.
api.crossref.orgblog.robotiq.comspectrum.ieee.orgodr.chalmers.se
Credit
B. Shimano; Victor Scheinman (Vicarm); Unimation
2021
OpenAI stops, short of data
In 2019 OpenAI showed a robot hand that solved a Rubik's Cube 20% to 60% of the time, guided by an AI model trained on the equivalent of 13,000 years of experience. In July 2021 its co-founder Wojciech Zaremba said the company had disbanded its robotics team, and that a lack of data "was holding us back".
LessonBy Zaremba's account, progress comes fastest where data is plentiful, and data from robots was not.
Credit
OpenAI
2023
Labs pool their robot data
In October 2023, 21 institutions released Open X-Embodiment, a combined set of recordings from 22 different robots covering 527 skills, and reported that a large model trained on it, RT-X, did better on several robots by drawing on the others' experience. In June 2024 a team from Stanford, UC Berkeley, Toyota Research Institute, Google DeepMind and others released OpenVLA, a vision-language-action model (an AI model that takes in camera images and a written instruction and puts out arm motions) with 7 billion parameters, trained on 970,000 robot episodes taken from the pooled set, with its weights and code open to all.
Credit
Open X-Embodiment Collaboration (293 listed authors); OpenVLA team (Moo Jin Kim, Karl Pertsch and others)
2025
ROS 2 replaces ROS 1
ROS, the Robot Operating System, is free, shared software for building robot programs, begun by Keenan Wyrobek and Eric Berger at Stanford and developed at Willow Garage, a private lab in California. The last version of the original ROS, called Noetic, reached the end of its support on 31 May 2025. Open Robotics, which looks after ROS, counted about 984 million ROS package downloads in 2025, with ROS 2 taking 91.2% of downloads in October 2025, and 1,579 companies on its private list of firms using ROS; it calls these figures a lower bound, because July's data was lost in a server move.
spectrum.ieee.orgdiscourse.openrobotics.orgdiscourse.openrobotics.org
Credit
Keenan Wyrobek and Eric Berger; Willow Garage; Open Robotics
2025
Open models, ranked blind
In February 2025 Physical Intelligence, a robot software company, released its π0 and π0-FAST models with their weights, the trained numbers that make up a model, and Hugging Face ported them into LeRobot, its open software for teaching robots (see the Note "No download"). RoboArena, described in June 2025 and first run at seven academic institutions, has evaluators at different sites run two models on the same kind of robot, an arm set up as in the DROID dataset, without knowing which model is which, on tasks they choose, and ranks models from which did better. On 18 November 2026 its public board listed 24 models; Physical Intelligence's π0.5 had been compared 745 times and ranked fifth, and none of the 24 had been online for comparisons in the previous 30 days.
Credit
Physical Intelligence; Hugging Face (Dana Aubakirova, Pablo Montalvo, Mustafa Shukor, Remi Cadene); RoboArena (Pranav Atreya, Karl Pertsch and 30 others)
Watch for
Five demonstrations, by NVIDIA's count
NVIDIA says GR00T 2, its robot model, succeeded 71% of the time on 50 new tasks after one demonstration of each, and 95% after five, which it puts at under 3 minutes of robot data. The tests ran on one two-armed robot with two hands that each move in 22 independent ways. NVIDIA chose the tasks and ran the trials, and on 18 November 2026 its page still read "Code Available in December" (see the Note "Five demonstrations per task").
Achieved whenThe code and weights are public and an outside team reports a similar success rate after five demonstrations, on tasks it chose itself.
Claimed by
NVIDIA, November 2026
Agents write robot programs
A team led from UC Berkeley's AUTOLab, with NVIDIA, built Graph-as-Policy, in which several AI coding agents assemble a robot program from a library of skills and rehearse it in a simulator, and says it clearly beats other systems. On the team's own benchmark, VABAR, programs written by the system with the Claude Opus 5 model scored below programs written by people in all 12 task-and-difficulty conditions on the Newton simulator: on the harder level of syringe packing, 12.1% against 63.0%. Claude Code, a coding agent, running the same model without the system scored 0% in 9 of the 12 (see the Note "Readable robot programs").
Achieved whenAn outside team runs agent-written programs on VABAR or on real robots and they match programs written by people.
Claimed by
UC Berkeley AUTOLab (Ken Goldberg) and NVIDIA, July to November 2026
Learning from its own attempts
Physical Intelligence says that training its π*0.6 model further on the robot's own attempts, good and bad, and on people's corrections, with a method it calls RECAP, more than doubled throughput and roughly halved the failure rate on some of its hardest tasks. The tasks it names are folding laundry in real homes, assembling boxes and making espresso drinks on a professional machine.
Achieved whenAn outside team applies the method to its own robot and task and publishes throughput and failure rates before and after.
Claimed by
Physical Intelligence, November 2025
Open
01
Demonstrations needed per task
An independent table of how a small arm's success rate rises with the number of demonstrations, for several open robot models. Ten tabletop tasks, such as stacking two cups or putting a pen in a box, are each taught with 1, 5, 20 and 50 demonstrations recorded by remote control; each model is fine-tuned on each amount and run 20 times per task.
The briefCredit · unclaimed02
Move a program between brands
A public test of AI models translating existing robot programs from one maker's language to another, for example ABB's RAPID to Universal Robots' URScript and back. Each translation is run in both makers' simulators to check that the tool follows the same path and switches its gripper and signals in the same order.
The briefCredit · unclaimed
Open 01 · the brief
Demonstrations needed per task
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An independent table of how a small arm's success rate rises with the number of demonstrations, for several open robot models. Ten tabletop tasks, such as stacking two cups or putting a pen in a box, are each taught with 1, 5, 20 and 50 demonstrations recorded by remote control; each model is fine-tuned on each amount and run 20 times per task.
Why it is still open
The few-demonstration figures for GR00T 2 come from NVIDIA's own test. Outside comparisons on the SO-101 arm held the amount of data fixed instead of varying it: Yi Yu and Xinchuan Qiu (June 2026) compared four models on four tasks, and SO-101 Bench (October 2026) trained every model on the same 3,629 simulated demonstrations. An April 2026 hobbyist trial on an SO-101 found 20 demonstrations not enough for one pick-and-place task; after 50, three models succeeded about 30% to 40% of the time, with a trial count given for only one of them (3 in 10).
research.nvidia.comarxiv.orghuggingface.cojustintiensmith.github.io
Built with what exists
An SO-101 arm kit, a hobby arm that AI Weekly puts at $120 to $150 in parts, with a second "leader" arm moved by hand to record demonstrations; LeRobot (see the Note "No download"); open models such as π0 and π0.5 through openpi or LeRobot, SmolVLA, and Runway's Praxis-1 (see the Note "Open robot model weights"); GR00T 2 once its code is released (see the Note "Five demonstrations per task"); one rented graphics processor for fine-tuning.
First step
Assemble an SO-101, record 50 demonstrations of one task, and fine-tune one model on the first 1, 5, 20 and 50 of them.
Done looks like
10 tasks, 4 demonstration counts and at least 3 models, with 20 filmed trials in each combination (2,400 trials), published as a table of success rates with every demonstration and video.
Open 02 · the brief
Move a program between brands
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of AI models translating existing robot programs from one maker's language to another, for example ABB's RAPID to Universal Robots' URScript and back. Each translation is run in both makers' simulators to check that the tool follows the same path and switches its gripper and signals in the same order.
Why it is still open
Published work has AI models write robot programs from plain-language instructions and checks them in simulation: the HARPA study (September 2026) for KRL, RAPID and URScript, and a 2026 Chalmers University thesis for RAPID in ABB's RobotStudio. None found translates an existing program from one brand's language to another. Offline programming software writes code for many brands from its own program, but RoboDK's moderator wrote in June 2024 that "it is not officially supported to import any type of robot program" into it, though some files, such as ABB's, may load.
Built with what exists
ABB RobotStudio, whose virtual controller runs ABB's robot software in simulation; Universal Robots' URSim simulator, published as a Docker image; an AI model such as Claude Fable 5.2 (see the Note "Three working devices"); a short script that logs the tool tip's position from each simulator.
First step
Collect 20 short RAPID example programs, run them in RobotStudio and save their tool paths as the answer key.
Done looks like
100 programs from vendor manuals and public examples, translated in each direction. Published: the share that runs without error, and the share whose tool path stays within 1 mm of the original at every taught point with its signals in the same order.