Meet The "Heavy Labor" Humanoid Robot Set To Revolutionize Construction

Japanese researchers at the Advanced Industrial Science and Technology (AIST) research center have developed a prototype humanoid robot, the HRP-5P, designed to autonomously perform heavy labor in hazardous environments. 

Standing just under 6-feet tall and clocking in at 222 lbs, the HRP-5P has "unsurpassed physical capabilities," according to, and is fitted with an array of sensors in order to fully assess its environment to perform various tasks. 

In one demonstration, the HRP-5P shows of its skills grabbing standard 77 lb. boards of drywall typically used in construction. 

    In order to complete the task, the robot must: 

    • Generate a 3-D map of the surrounding environment, detect objects, and approach the workbench.
    • Lean against the workbench, slide one of the stacked gypsum boards to separate it, and then lift it.
    • While recognizing the surrounding environment, carry the gypsum board to the wall.
    • Lower the gypsum board and stand it against the wall.
    • Using high-precision AR markers, recognize and pick up a tool.
    • Holding a furring strip to keep HRP-5P itself steady, screw the gypsum board into the wall.

    Aside from construction, AIST's robot has compelling applications at aircraft facilities, shipyards or any other environment in which heavy things need to be lifted or manipulated - particularly in hazardous environments. 

    AIST collaborated with several private companies in the development of the HRP series, including Kawada Robotics, which has assisted in the design of several prototypes leading up to the 5P. HRP-2, for example, was able to walk, lie down, stand up, walk on narrow paths, and other navigational maneuvers. Its successor, the HRP-3 was able to walk on slippery surfaces and tighten bridge bolts via remote control. 

    The 5P marked a significant leap over its predecessors - employing technology from Honda Motor Co, and the New Energy and Industrial Technology Development Organization (NEDO), the latest and greatest from AIST can work in "unstructured environments," and excels at "targeting full-body motion planning based on environmental model acquisition that enables humanoid robots to adapt to unknown environments." 

    More specs via AIST:

    • At a height of 182 cm and weight of 101 kg, HRP-5P has a body with a total of 37 degrees of freedom: two in its neck, three in its waist, eight in its arms, six in its legs, and two in its hands. Except for the hands, this represents the most freedom of movement in the HRP series to date. Compared to the revised version of HRP-2, adding one degree of freedom to the waist and one to the base of the arms has enabled operations more closely resembling human motion. Accordingly, using both arms, HRP-5P can handle large objects such as gypsum boards (1820 × 910 × 10 mm, approx. 11 kg) or plywood panels (1800 × 900 × 12 mm, approx. 13 kg).
    • To emulate human motion by the robot without as many degrees of freedom as people, the researchers ensured a wider movable range of joints in the hip and waist areas, where multiple joints are concentrated. For example, hip joints that flex and extend the legs have a range of motion of 140° in humans and 202° in HRP-5P (Fig. 1), and waist joints that turn the upper body have a range of motion of 80° in humans and 300° in HRP-5P. This enables work by the robot in a variety of postures, such as when deeply crouched with the upper body twisted.
    • Joint torque and speed were approximately doubled on average relative to the revised HRP-2, by employing high-output motors, adding cooling to the drive mechanism, and adopting a joint drive system with certain joints featuring multiple motors. As a result, the robot can do work involving heavy loads, such as lifting a gypsum board from a stack. (Each arm of HRP-5P, extended horizontally, can bear a weight of 2.9 kg, compared to 1.3 kg for the revised version of HRP-2 and 0.9 kg for HRP-4.)
    • Using head-mounted sensors, the robot constantly acquires 3-D measurements of the surrounding environment (at a frequency of 0.3 Hz). Even if the field of view is blocked by objects used in work, stored and updated measurement results enable execution of the walking plan while carrying a panel or correction of walking when the feet slip. (Fig. 2).
    • Learning involves a convolutional neural network using a newly constructed image database of work objects. The robot can detect ten types of 2-D object regions at a high precision of 90 % or more even against low-contrast backgrounds or under dim lighting (Fig. 3).
    • It was possible to build a highly reliable robot system and maintain the quality of large-scale software (with approx. 250,000 lines of code) by arranging a virtual test environment for the  in the Choreonoid robot simulator and monitoring software regression for 24 hours.