Controlling Assistive Robots with Learned Latent Actions

Dylan P. Losey, Krishnan Srinivasan, Ajay Mandlekar, Animesh Garg, Dorsa Sadigh

★ Must read September 2019

TL;DR They train a state-conditioned VAE on demonstrations so that a 1- or 2-DoF joystick input decodes into a full action for a 7-DoF assistive arm. In an eating study users reached the intended morsel in 44 of 50 trials, more often than with shared autonomy, and in a cooking study they finished faster with less joystick input than with end-effector control.

What they do

Assistive arms have more degrees of freedom than a user can coordinate with a joystick. The paper learns latent actions: a low-DoF embedding of demonstrated robot actions, so that a 1- or 2-DoF joystick input is decoded into a full high-DoF action. It defines what makes latent actions user-friendly (controllability, consistency, scaling), tests which autoencoders provide them in simulation, and runs two user studies on a 7-DoF Fetch arm.

How it works

Autoencoder. An encoder maps a demonstrated state-action pair to a latent z, and a decoder reconstructs the action. The loss is reconstruction error, plus a KL term towards a unit normal for the VAE variants.
State conditioning. The decoder takes both z and the current state (cAE, cVAE), so the same joystick push can mean “pour” when the robot holds a cup. Baselines are PCA, AE and VAE without state input.
Control. The learned latent axes are rotated by hand to line up with the joystick axes. At run time the joystick input is z and the decoder outputs the robot action.

Results

Simulation (planar 5-joint arms): on Sine, cAE and cVAE reach 1.37% and 3.74% of PCA’s reconstruction loss, versus 98.0% and 100% for AE and VAE; on Rotate, 0.65% and 0.84%. On Circle the two latent directions are 72 and 74 degrees apart for cAE and cVAE, versus 27 and 34 for AE and VAE (90 is ideal).
Eating (10 users): with cVAE, users reached their intended morsel in 44 of 50 trials, a higher success rate than the shared-autonomy baselines. Completion time, joystick input and trajectory length were all significantly lower than with High Assist.
Cooking (11 users, within-subjects; cVAE trained on under 7 minutes of kinesthetic demonstrations): completion time and joystick input dropped compared with end-effector teleoperation (p < .001), and users rated cVAE as less effort, easier and more natural.

Limitations

Outside the states covered by demonstrations the latent actions became erratic: one participant turned the arm upside-down while pouring flour and could not guide it back. Users did not clearly prefer latent actions, saying they enjoyed the freedom of end-effector control. In the eating study participants used only cVAE; the baselines come from the HARMONIC dataset.

Videos

Figures

From Figure 1: A user teleoperates the arm with a 2-DoF latent action trained on pouring demonstrations: z1 moves the... — From Figure 1: A user teleoperates the arm with a 2-DoF latent action trained on pouring demonstrations: z1 moves the cup level with the table and z2 tilts it. Right: a conditional autoencoder encodes state and action into z, and the decoder turns z plus the current state back into a high-DoF action.

From Figure 2: Sine task. (A) Test loss of VAE and cVAE, normalized by the PCA loss. (B) Effect of the cVAE latent ac... — From Figure 2: Sine task. (A) Test loss of VAE and cVAE, normalized by the PCA loss. (B) Effect of the cVAE latent action z at three states along the sine wave. (C) Rollouts of VAE and cVAE from the same start under a constant input z = +1. (D) End-effector paths over several rollouts.

From Figure 6: Eating study. A participant guides the arm to their chosen morsel with a 2-DoF joystick (left); succes... — From Figure 6: Eating study. A participant guides the arm to their chosen morsel with a 2-DoF joystick (left); success rate of cVAE against the No, Low, High and Full Assist shared-autonomy baselines from the HARMONIC dataset (right).

From Figure 9: Cooking study in three sub-tasks: pour eggs and recycle the container, pour flour and return it to the... — From Figure 9: Cooking study in three sub-tasks: pour eggs and recycle the container, pour flour and return it to the shelf, add an apple and stir (top row). Example robot trajectories under End-Effector teleoperation (middle row) and with cVAE latent actions (bottom row).

Abstract

Assistive robotic arms enable users with physical disabilities to perform everyday tasks without relying on a caregiver. Unfortunately, the very dexterity that makes these arms useful also makes them challenging to teleoperate: the robot has more degrees-of-freedom than the human can directly coordinate with a handheld joystick. Our insight is that we can make assistive robots easier for humans to control by leveraging latent actions. Latent actions provide a low-dimensional embedding of high-dimensional robot behavior: for example, one latent dimension might guide the assistive arm along a pouring motion. In this paper, we design a teleoperation algorithm for assistive robots that learns latent actions from task demonstrations. We formulate the controllability, consistency, and scaling properties that user-friendly latent actions should have, and evaluate how different low-dimensional embeddings capture these properties. Finally, we conduct two user studies on a robotic arm to compare our latent action approach to both state-of-the-art shared autonomy baselines and a teleoperation strategy currently used by assistive arms. Participants completed assistive eating and cooking tasks more efficiently when leveraging our latent actions, and also subjectively reported that latent actions made the task easier to perform. The video accompanying this paper can be found at: https://youtu.be/wjnhrzugBj4.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong