RLで新しいこと試すとき、初期値として適当なポリシーから追加学習するほうが収束が手っ取り早いので
都度最新の実機搭載ポリシーを基準に追加学習をさせているのだが
気づいたら現在の最新ポリシーの総iterが100000超えていたのでびっくり
2
297
NWK/野分機械 retweeted
❗ JUST IN: Boston Dynamics just put wheels on Atlas. And the result is pretty incredible! 🤖
Instead of walking, the humanoid can now roll, turn, slide sideways, and even perform a 360-degree jump.
But what's really interesting is why the engineers built it.
This isn't a new version of Atlas intended for customers. It's an experiment to understand how far they can push the robot's motors and control systems.
And the hardware modification was surprisingly simple.
The team replaced the feet with off-the-shelf scooter wheels, using just one custom adapter design to connect them to the existing ankle actuators.
The real challenge was getting the robot to control them.
Atlas uses reinforcement learning, where a controller learns movements through millions of simulated experiences before being deployed on the physical robot.
But there's a problem.
Simulated motors don't automatically behave like real ones.
If the simulation assumes a motor can deliver more torque or respond faster than it actually can, the robot might learn movements that are physically impossible.
So the team carefully matched the simulated motors to their real-world behavior, including how they perform under high loads and at different speeds.
With a few adjustments to the training setup, Atlas learned to roll smoothly while maintaining balance.
And some pretty fascinating behaviors emerged.
When performing sideways dashes, for example, the controller learned to steer the wheels in ways that resemble the sliding movements of a human.
What's particularly interesting is that the lessons from this experiment are already being applied to the regular walking version of Atlas.
I love seeing projects like this because they show that making robots more capable isn't always about adding more complicated hardware.
Sometimes changing one component is enough to discover entirely new ways for a robot to move.
Credit: @BostonDynamics
👉 Follow for more if you are interested in robotics!
歩行安定性向上やカメラ搭載等色々とアプデして望んだ本イベント
操縦の没入感を高めるためのカメラ映像配信が、まさかのバグで動かず…!
とはいえロボ自体は終日良く動いてくれたし、データも取れたので全体としてはプラス
おつかれさまでした!
#NT湘南
Zero 2 W移行後、DELTA-VのRL歩行が不安定に。制御周期は守れており、DRでも改善せず原因不明だった。
実は、前周期のactionを周期冒頭に送るはずが、推論終了後に送っていた潜在バグが発覚。CM5では3msだった推論時間が12msに増えたことで顕在化した模様。
修正したらかなり安定した。よかった。
15
657


















