Fake data helps robots learn the ropes faster: But how?

[Sum­ma­ry: A new approach could cut learn­ing time for robots that work with soft objects like ropes and fab­rics. In sim­u­la­tions, the expand­ed train­ing data set dou­bled suc­cess rate of a robot loop­ing a rope around an engine block. Using only the ini­tial train­ing data, the sim­u­lat­ed robot hooked the rope around the engine block 48% of the time. After train­ing on the aug­ment­ed data set, the robot suc­ceed­ed 70%.]


A new approach could cut learn­ing time for robots that work with soft objects like ropes and fabrics.

It was devel­oped by robot­ics sci­en­tists at the Uni­ver­si­ty of Michi­gan and might min­i­mize the time it takes to learn new things and set­tings from a week or two to a few hours.

For their false data, they focused on the three fea­tures. It has to be cred­i­ble, diverse, and relevant.

The sim­u­lat­ed robot suc­cess­ful­ly wrapped the rope around the engine block 48% of the time using only the basic train­ing data.


A new method enhances train­ing data sets for robots that func­tion with soft objects like ropes and fab­rics, or in dense set­tings, advanc­ing toward devel­op­ing robots that can learn on the fly like peo­ple do.

Robot­ics sci­en­tists at the Uni­ver­si­ty of Michi­gan devel­oped it and might min­i­mize the time it takes to learn new things and set­tings from a week or two to a few hours.

In sim­u­la­tions, the larg­er train­ing data set more than dou­bled the suc­cess rate of a robot loop­ing a rope around an engine block and improved it by more than 40% from that of a phys­i­cal robot exe­cut­ing the same task.

That is one of the activ­i­ties a robot mechan­ic would need to gain enough skill at. How­ev­er, accord­ing to Dmit­ry Beren­son, U‑M asso­ciate pro­fes­sor of robot­ics and senior author of a paper pre­sent­ed today at Robot­ics: Sci­ence and Sys­tems in New York City, learn­ing how to manip­u­late each unfa­mil­iar hose or belt would require tremen­dous large amounts of data, like­ly col­lect­ed for days or weeks.

Dur­ing that peri­od, the robot would exper­i­ment with the hose, extend­ing it, con­nect­ing its ends, loop­ing it around items, and so on, until it was aware of all the pos­si­ble move­ments the hose could make.

Read:

Beren­son said “If the robot needs to play with the hose for a long time before being able to install it, that’s not going to work for many applications.”

Cer­tain­ly, a robot col­league that need­ed that much time would prob­a­bly not be well regard­ed by human mechan­ics. Thus, Beren­son and Peter Mitra­no, a robot­ics PhD stu­dent, altered an opti­miza­tion algo­rithm to assist com­put­ers to make some of the gen­er­al­iza­tions that peo­ple do, such as pro­ject­ing how dynam­ics seen in one instance would recur in others.

In one occa­sion, the robot moved cylin­ders across a packed floor. The cylin­der some­times did­n’t make con­tact with any­thing, but oth­er times it did and the oth­er cylin­ders shift­ed as a result.

You can replay the motion any­where on the table where the tra­jec­to­ry does not dri­ve it into oth­er cylin­ders if the cylin­der did­n’t hit with any­thing. A human would com­pre­hend this, but a robot would need to acquire that data. 

And Mitra­no and Beren­son’s pro­gram can also pro­vide vari­a­tions on the out­put of that first exper­i­ment that aid the robot in the same way rather than under­tak­ing time-con­sum­ing exper­i­ments.

For their false data, they focused on the three fea­tures. It has to be cred­i­ble, diverse, and rel­e­vant. Data on the floor, for instance, is mean­ing­less if you you’re focus­ing only on the robot mov­ing the cylin­ders on the table. 

On the oth­er hand, the data must be var­ied; all parts of the table and all view­points must be evaluated.

“If you max­i­mize the diver­si­ty of the data, it won’t be rel­e­vant enough. But if you max­i­mize rel­e­vance, it won’t have enough diver­si­ty,” Mitra­no said, adding that both are important.

Related Stories:

The data must also be accu­rate. For illus­tra­tion, any sim­u­la­tions in which two cylin­ders occu­py the very same space would need to be labeled as invalid in order for the robot to know that this won’t happen.

Mitra­no and Beren­son broad­ened the data set for the rope sim­u­la­tion and exper­i­ment by extend­ing the rope’s posi­tion to addi­tion­al loca­tions in a dig­i­tal form of a real world sit­u­a­tion, assum­ing the rope would act the same as it did in the ini­tial case. 

The sim­u­lat­ed robot suc­cess­ful­ly wrapped the rope around the engine block 48% of the time using only the basic train­ing data. The robot has 70% rate of suc­cess after train­ing on the expand­ed data set.

Hav­ing the robot expand each effort in this way near­ly dou­bles its suc­cess rate over the course of 30 attempts, with 13 suc­cess­ful attempts as com­pared to sev­en, accord­ing to an exper­i­ment study­ing on-the-fly learn­ing with a real robot.

This find­ing will be a big step for robot­ics research and human robot inter­ac­tion. This change will be help­ful to the devel­op­ment of robots that can learn on the fly like humans do.

In addi­tion, sci­en­tists have also pre­dict­ed that this research will pave the way for bet­ter under­stand­ing on how to use non-tra­di­tion­al meth­ods in robot­ics such as data aug­men­ta­tion approach­es to get bet­ter results.

The Toy­ota Research Insti­tute, the Office of Naval Research, the Nation­al Sci­ence Foun­da­tion, and IIS-1750489 and IIS-2113401 funds all pro­vid­ed sup­port for this research.

Leave a Reply

Your email address will not be published.

Join our NewsletterDaily Glimple of Future

Our blog, "Daily Glimpse of Future", strives to make the future much clearer than it is today. Join our newsletter for free now.