Raibo, el robot surcoreano que hace parkour

xataka

Vídeo: Robotics & Artificial Intelligence Lab

Transcript

00:00This research proposes a hierarchical navigation module capable of high-speed legged locomotion

00:04even in complex and discreet terrain distribution environments

00:07and performing extreme parkour maneuvers such as running on walls or jumping over large gap.

00:16The proposed navigation module robustly operates in complex environments

00:21composed of various types of obstacles.

00:23We have structured the entire pipeline into three modules.

00:26The planner module, the tracker module, and the target updater.

00:30The planner module generates a feasible foothold from the map.

00:33The target updater changes the target foothold for the robot to step on.

00:37The tracker module controls the robot to accurately step on the given target foothold.

00:42We first trained the tracker module to achieve high performance during the training stage.

00:46The tracker module was trained using reinforcement learning on the Razum simulator.

00:50We employed a generative model called Map Generator, which evolves adversarially with the tracker module

00:55to provide desired difficulty environments.

00:56At deployment stage, we can observe the operation of each module on the real experiment.

01:02While the tracker module is operating, if the robot sufficiently steps on the target foothold,

01:06target updater updates the foothold target.

01:09When the front target is updated, it activates the planner module.

01:12The planner module runs simultaneously on a detached thread.

01:15While the tracker module controls the robot to precisely step on next target,

01:18planner module generates a feasible and safe foothold plan.

01:21To reduce time complexity in planning, the rear foot steps where the front foot has stepped.

01:26Planner draws multiple foothold samples through random sampling and rejects risky and difficult

01:31footholds using a consecutive filtering structures.

01:34Subsequently, the candidate foothold plans are evaluated through rollout in the physics simulation,

01:39and the one with the lowest cost is selected and provided to tracker as a next target.

01:43The planning module performs sampling and rollouts in parallel across eight different threads,

01:49separate from the tracker module.

01:51The real robot controlled by the tracker module is depicted in gray,

01:54while the rollout process is shown in blue.

01:57The final goal position is represented by a light green cube.

02:00The robust navigation module performs well even on maps with various unstructured obstacles.

02:09These maps include maps with pillars of various shapes and heights arranged sequentially,

02:13maps with impassable pillars or pits in discrete terrains,

02:16maps with 1.1-meter patches tilted randomly up to 30 degrees,

02:19maps with steps of varying heights and lengths,

02:22maps with consecutive ramp pillars of various heights and angles,

02:25and maps requiring descent or ascent of high ledges.

02:27The navigation module operates robustly even when the goal position is suddenly changed.

02:33It can traverse approximately 90 meters through maps with mixed obstacles

02:36at an average speed of 2.7 meter per second

02:39and maximum speed of 3.63 meter per second without a single failure.

02:43This experiment demonstrates the powerful performance of the tracker module,

02:47a map requiring ascent and descent of 0.6 meter obstacle,

02:52a map requiring side flutter,

02:54a map requiring jump-across-gap up to 1.3 meters wide,

03:00a map requiring traversal of double ramps facing in opposite directions,

03:10a map consisting of triple consecutive ramps,

03:19a map requiring running on a single wall.

03:28The following video demonstrates the performance of the entire pipeline in a complex environment.

03:45The scenario involves a map with stairs and various obstacles,

03:48where the goal is positioned on the right.

03:50As before, the scenario involves a map with stairs and various obstacles,

04:00but this time the goal is positioned on the left.

04:02This scenario involves a map with various sized boxes and ramps,

04:17where the cost for avoiding height changes is low.

04:19This scenario involves a map with various sized boxes and ramps,

04:32where the cost for avoiding height changes is high.

04:34This scenario involves a map with boxes and slanted surfaces requiring abrupt directional changes,

04:53with gaps up to 1.3 meters wide.

04:55This scenario involves a map with boxes and slanted surfaces requiring abrupt directional changes,

04:56with gaps up to 1.3 meters wide.

04:59soilما above 2 Nevertheless,

05:01being made carefully around 3 e.8 meters wide and

05:14more powerful91 meters beyond the left to place.

05:18A map with facts,

05:21viaje to 2,0 meters wide and continue to form the dangers of a distributed position.

Categoría

Transcripción

Recomendada