Installation and hands-on records and adoption notes from the field. The numbers come from what we installed and measured ourselves.
The educational AMR robot arm we loaded into Isaac Sim was entirely white, so we tried to color each part. We created a material and bound it, but for six attempts the robot stayed white. The cause was neither the lighting nor the material type but "where the material was bound." This is a record of ruling out one cause after another behind the same symptom.
Read →
We used reinforcement learning to teach only one step, approaching the block, to an educational AMR robot arm. Of 11 trainings, four collapsed on reward loopholes such as "standing still," and one that looked good by training score never actually arrived. The final policy arrived in 199 of 200 simulated runs but fell short of the target time.
Read →
We put an approach policy trained with reinforcement learning in a simulated environment onto a physical AMR robot arm and compared it with a hand-written rule, three runs each from the same position. Approach time was 13.2 s versus 19.4 s, and the "lower the camera two steps at once" behavior learned in simulation appeared on the physical robot at about the same distance. The result comes with conditions attached, though - one position, three runs each, and a low battery.
Read →
An educational AMR robot arm learned to stand right in front of a block in a simple simulation without a physics engine, in a short time. It succeeded in all 100 of 100 conditions not used in training, but we first asked what that 100% means. When we moved to a 3D physics engine, the robot flew off once, and only after a fix did it succeed in one condition.
Read →
We mounted a lidar and an IMU on the educational AMR robot arm standing in the virtual world. We put walls on all four sides and checked whether the lidar points land on the walls. Read the scan in the opposite direction and only 42.8% match. The IMU stayed at 0 because of one thing, the order in which it was created. We also describe how we handled a lidar whose real specification we could not obtain.
Read →
Before starting reinforcement learning, we built the world the simulated robot would practice in first. A practice arena that places obstacles in a different spot every time, an observation that reduces thousands of lidar points to a few dozen numbers, and a pose that bends the arm so the camera looks at the floor. We have not trained yet; this is a record of checking that the arena runs without breaking.
Read →
While teaching the arm of an educational AMR robot arm to grasp a 3 cm cube with reinforcement learning, we met a reward loophole twice, where only the score rose and the cube was not lifted. After we fixed the reward, the policy passed 98 of 100 attempts in the calculation-only practice environment, but in a physics simulation that also computes contact, it failed to lift the cube in all 4 attempts. We wrote down the process as it happened.
Read →
We put an educational AMR robot arm into an Isaac Sim virtual world and turned gravity on. Within one second it was thrown out of the scene. We switched off suspect settings one by one to rule out causes, and with merging of parts and automatic inertia calculation both turned on, the robot held its place for 50 seconds. This post also covers how the gripper broke the calculation and the problem we have not solved.
Read →