How much practice does a factory robot need?

Two hours of guidance for a foam part

Eugenio Chisari and colleagues at Mimic Robotics and BMW Group · November 11 preprint · Physical robot tests in a controlled cell

Robots that learn by imitation need demonstrations. A person operates the robot through a task while its cameras and motors are recorded. In a car plant every station has its own parts and fixtures, so the hours of demonstration each new task requires go a long way toward deciding whether the approach is affordable. Mimic Robotics and BMW chose tasks by watching manual work at automotive sites, then recorded about 296 hours across 31 of them. Roughly a quarter came from people wearing camera gloves, with no robot involved.

Their system, FLUX-mimic, starts from a model trained to predict video, which was also shown all of that factory footage. A smaller part then learns to turn those predictions into arm and hand movements. On a pair of arms with multi-fingered hands it completed four tasks at an average of 82.8%: mounting a control unit, inserting it into a fixture, sorting foam strut covers into slots, and clipping a cable to a door panel. A widely used general-purpose robot model called π0.5 averaged 27.6%, and a smaller policy trained separately for each task averaged 52.6%. With one hour of demonstrations for the foam task, FLUX-mimic succeeded in 45% of 40 attempts, against 7.5% and 2.5%. At eight hours it reached 95%.

Each task got 18 or 20 attempts, in a dedicated cell with uniform lighting and background. The cable task succeeded 61.1% of the time. On a second robot with simple two-finger grippers, FLUX-mimic and π0.5 tied at 60% across two tasks. Video of the test tasks was part of the model’s earlier training, so the hours counted here cover only demonstrations with recorded robot movements. The authors work at the two companies, and the paper links no code or data.

• Paper ↗

A €3,200 robot cleans a toilet seat, most of the time

Sven Schreiber and colleagues at the University of Hamburg, Hon Hai Research Institute and Northeastern University in China · November 11 preprint · Physical robot tests

Many well-known robot results come from machines that cost tens of thousands of euros. This team mounted two hobbyist SO-101 arms on a TurtleBot 4, a small wheeled base common in teaching labs, for about €3,200 in total. The job was to fetch a wet sponge from a table, drive to a toilet, scrub a food-coloring stain with the rough side, flip the sponge using both arms, wipe with the soft side, and put the sponge back.

The team recorded 250 demonstrations and used them to adapt four openly available robot models. One, called π0, removed the stain completely in 12 of 20 attempts and left only a faint trace in two more. When a person was allowed to reposition a stuck robot once, 18 of 20 runs finished. The released code for its successor, π0.5, managed 5 of 20. Two add-ons meant to help a robot keep track of its place in a long task both lowered the scores. The two remaining models never succeeded.

This is one toilet in one room under fixed lighting, with 20 attempts per policy except GR00T, which the team stopped testing after five failures. The public π0.5 code lacks parts of the system its makers described, so the comparison is with what anyone can download and not with the full model. The robot ran on external power, with a separate computer choosing its movements. The project page says code and data will be released if the paper is accepted.

• Paper ↗
• Project page ↗

A ladder of chores, from easy to nearly impossible

Chongkai Gao, Lin Shao and colleagues at the National University of Singapore and eight other universities · November 12 preprint · Simulation, with eight physical spot checks

Many robot tests are collections of similar pick-and-place chores, so a high score says little about how hard a task a system can handle. ManiLadder sorts 112 simulated tasks into four levels. They cover rigid objects and soft ones such as rope, cloth and clay, one arm or two, and grippers or multi-fingered hands. The authors measured difficulty. They trained one standard policy on 20 demonstrations of each task and adjusted the task until its success rate fell in a target band: 60% or more for level one, 4% or less for level four.

No approach led everywhere. π0.5 beat the small standard policy on some groups of tasks and lost to it on others. A coding agent that wrote its own control programs did best on several harder rigid-object groups and scored zero on one soft-object group. A single π0.5 policy trained on all 112 tasks together averaged 8.6%, against 33.5% for separate policies. On sixteen level-two tasks, 200 demonstrations per task still left the average below 32%. That is ten times the standard amount.

The levels are tied to one policy and one amount of data. They describe what is hard for today’s imitation methods, which may differ from what is hard for other approaches. The coding agent got extra practice runs, which makes its comparison uneven. On eight tasks rebuilt with a physical arm, at 50 trials each, success fell level by level much as it did in simulation. The first author has published demonstration trajectories, camera recordings and normalization statistics on Hugging Face.

• Paper ↗
• Demonstration data ↗

Agile Robots raises $500 million

Agile Robots · November 12 announcement · Company-reported figures

Munich-based Agile Robots said SoftBank led a $500 million funding round that values the company at $8.5 billion, with NVIDIA, Hitachi and T.Capital also investing. The valuation is the company’s own figure. It says the money will go to expanding in the United States and raising production, including in Germany.

Agile reports 20,000 industrial robot systems deployed with automotive, electronics and logistics customers. Its subsidiaries include Franka Robotics, which makes the arms used in the BMW study above. Its site near Munich is built to make up to 10,000 humanoids a year, but the announcement gives no number of humanoids in use.

• Announcement ↗