End-to-End Neural Low-Level Control of an Industrial-Grade 6D Magnetic Levitation System

Philipp Hartmann, Jannick Stranghöner, and Klaus Neumann

[Code on Github] (released upon publication)

[See Paper]

[Our research group] (released upon publication)

Data Generation utilizing Proprietary Controller

The proprietary controller is utilized to track augmented 6D lift-off trajectories at randomly sampled poses. The augmentations are realized through reference pose jumps of up to 0.8mm or 0.8° per axis. This data is used to train a GRU-based neural controller via behavior cloning. This neural controller is evaluated after this section.

lift-off_muted.mov

Evaluation of Neural Controller

Here, we evaluate the learned GRU-based neural controller. It is trained solely on 10000 lift-off trajectories (one is illustrated above).

1) We first showcase (a) qualitative robustness, (b) stability to discontinuous reference pose jumps and (c) stability to continuous reference pose oscillations.

2) Next, we demonstrate generalization to (a) large circles (30mm radius) with different velocities/frequencies and (b) a random trajectory over the whole tile with different velocities.

(1a) Qualitative Robustness of behavior-cloned Neural Controller

Here we demonstrate qualitative robustness of the behavior-cones neural controller. Despite that the neural controller has never transportation of weights or force perturbations, it is capable of carrying some weights and remains

small weight (30g)

medium weight (125g)

robustness-smallweight_muted.mov

robustness-mediumweight_muted.mov

large weight (325g)

manual forces

robustness-largeweight_muted.mov

robustness-manual_muted.mov

(1b) Response to Reference Pose Steps of behavior-cloned Neural Controller

For each reference pose jump, the mover is at a randomly sampled starting position. Then a discontinuous reference pose step is performed in which the reference pose is abruptly changed by up to 0.8mm or 0.8deg per axis. After that, the controllers have time to recover from that discontinuous jump.

reference-pose-jumps_muted.mov

(1c) Response to Oscillation (Bode Plot)

Here, we utilize oscillations with as set of frequencies from 0.1hz up to 25.6hz and an amplitude of 0.5mm or 0.5°. This is done independet on each axis to empirically obtain Bode plots to demonstrate stability to dynamic continuous perturbations.

bode_muted.mov

(2a) Circular Trajectory (large)

Here, we track a circular trajectory with radius of 30mm. We utilize different frequencies of 0.25hz, 0.5hz, 1hz, and 2hz.

r=30mm, f=0.25hz

r=30mm, f=0.5hz

circle-large-slow_muted.mov

circle-large-middle_muted.mov

r=30mm, f=1hz

r=30mm, f=2hz

circle-large-fast_muted.mov

circle-large-ultrafast_muted.mov

(2b) Random Trajectory Tracking

Here, we track a random trajectory with three different velocities.

random-traj-slow_muted.mov

random-traj-middle_muted.mov

random-traj-fast_muted.mov

Overview of random trajectory

Overview of root mean squared error

Future Work (coming soon)

Currently working on a physics-based simulation environments, to simulate the whole system.

This allows everyone to participate in this exciting environment.

Here, the neural controller operates in simulation:

gpu_sim_final.mp4

Page updated

Google Sites

Report abuse