Underwater waste — removal by robots

Col­lect­ing trash in oceans and seas is very time-con­sum­ing and expen­sive. A team from the Tech­ni­cal Uni­ver­si­ty of Munich (TUM) is devel­op­ing a robot­ic sys­tem in a Euro­pean coop­er­a­tion project that can locate and col­lect waste under­wa­ter. Machine learn­ing meth­ods are being used in the process.

There are cur­rent­ly between 26 and 66 mil­lion tons of plas­tic waste in our seas and oceans. Most of this is on the ocean floor. This pos­es a huge threat to marine life and the eco­log­i­cal bal­ance of the oceans.

But clean­ing up the waters is com­plex, expen­sive and often dan­ger­ous, because divers usu­al­ly have to be deployed. In addi­tion, clean­ing oper­a­tions are usu­al­ly con­cen­trat­ed on the water sur­face. A team from TUM has joined forces with eight Euro­pean part­ner insti­tutes in the “Sea­Clear” project to devel­op a robot sys­tem that can also col­lect trash underwater.

Four robots act together

The sys­tem is made up of four indi­vid­ual robot com­po­nents: An autonomous­ly oper­at­ing robot­ic boat per­forms an ini­tial scan of the seafloor, locat­ing major trash accu­mu­la­tions. Then, an obser­va­tion robot is launched into the water to detect trash at depth while pro­vid­ing addi­tion­al infor­ma­tion, such as close-up images of the seafloor, to the computers.

When the water is clear and vis­i­bil­i­ty is good, an aer­i­al drone also ensures that fur­ther trash is detect­ed in the water. Using all this infor­ma­tion, a vir­tu­al map is gen­er­at­ed. A col­lec­tion robot then trav­els to cer­tain points on the map and col­lects the trash. In the process, larg­er pieces are removed using a grab­ber in a bas­ket con­nect­ed to the ship.

The challenge of currents

“Devel­op­ing autonomous robots for use under­wa­ter pos­es a very spe­cial chal­lenge,” says Dr. Ste­fan Sos­nows­ki, tech­ni­cal direc­tor of the Sea­Clear project at the Chair of Infor­ma­tion Tech­nol­o­gy Con­trol at TUM. This is because, unlike on land, very spe­cial con­di­tions pre­vail in the water. “Once a piece of trash has been iden­ti­fied and locat­ed, the robot must first move close to it. In doing so, it can some­times encounter strong cur­rents, against which it must pre­vail. Get­ting that right is TUM’s job in the Sea­Clear project.”

Efficient machine learning

To do this, the team is using machine learn­ing meth­ods. An arti­fi­cial intel­li­gence (AI) cal­cu­lates and learns when and under which con­di­tions the robot moves in a cer­tain way. This allows accu­rate pre­dic­tions to be made about its behavior.

“Anoth­er chal­lenge is that we don’t have the com­put­ing pow­er we’re used to on land” says Prof. San­dra Hirche, head of the depart­ment and Sea­Clear project coor­di­na­tor. “There is no con­nec­tion to large data cen­ters with high-per­for­mance com­put­ers. The algo­rithms we devel­op there­fore have to be as effi­cient and resource-effi­cient as pos­si­ble. That’s why we’ve been work­ing for some time on meth­ods with high “sam­pling effi­cien­cy” that can make the best pos­si­ble pre­dic­tions with as lit­tle data as pos­si­ble. Infor­ma­tion that is not need­ed is sim­ply for­got­ten by the AI.”

90% success rate

Once ful­ly oper­a­tional, the Sea­Clear sys­tem is expect­ed to clas­si­fy under­wa­ter debris at a pre­dict­ed rate of 80% and suc­cess­ful­ly col­lect it 90% of the time — com­pa­ra­ble to the suc­cess of using divers. Ini­tial tri­als of the pro­to­type were con­duct­ed in Octo­ber 2021 under clear water and vis­i­bil­i­ty con­di­tions in Dubrovnik, Croa­t­ia. There are to be fur­ther tri­als in the port of Ham­burg in May 2022.