DayDreamer: World Models for Physical Robot Learning

Philipp Wu, Alejandro Escontrela, Danijar Hafner, Ken Goldberg, Pieter Abbeel

Seminal June 2022

TL;DR Runs the Dreamer world-model RL algorithm online on four real robots, with no simulator and one set of hyperparameters. A Unitree A1 learns to roll over, stand up and walk in about 1 hour, and two arms learn visual pick-and-place from sparse rewards at close to human pick rates.

What they do

Deep RL needs too much trial and error for real hardware, so most robot learning goes through a simulator. This paper tests whether Dreamer, a world-model RL algorithm shown on video games and simulated tasks, can learn online on physical robots. It applies Dreamer unchanged, with identical hyperparameters, to an A1 quadruped, a UR5 arm, an XArm and a Sphero Ollie wheeled robot.

How it works

World model. A Recurrent State-Space Model: an encoder fuses all sensors (joint readings, RGB, depth) into stochastic latents, a recurrent dynamics model predicts their sequence, a decoder reconstructs the inputs and a reward head predicts rewards. It predicts future representations rather than future images.
Behavior learning in imagination. An actor and critic train on 16-step latent rollouts with batch sizes of 16K on one GPU. The critic regresses lambda-returns; the actor uses reparameterization gradients for continuous actions and Reinforce for discrete ones.
Asynchronous setup. Built on DreamerV2, a learner thread trains continuously while an actor thread computes actions, which matters for the quadruped’s 20 Hz control.

Results

A1 walking: rolls off its back within 5 minutes, stands up 20 minutes later and walks with a pronking gait after about 1 hour, without resets. SAC learns to roll over but never stands or walks. When pushed, it adapts within 10 minutes.
UR5 pick and place (3 balls, sparse rewards, discrete actions): 2.5 objects per minute after 8 hours, approaching human operators. Rainbow and PPO only learn to grasp and drop in the same bin.
XArm pick and place: 3.1 objects per minute in 10 hours, comparable to humans; Rainbow fails. After sunrise lighting collapses performance, it recovers in about 5 hours.
Sphero navigation from images: reaches the goal within 2 hours (average distance 0.15 of the area size), similar to DrQv2.

Limitations

Hours of training wear the robots and can need human intervention or repair. The authors say longer training is needed to find the limits of Dreamer and the baselines. The tasks are simplified: the A1 is manually repositioned at the edge of the training area, the XArm’s object is tied to the gripper with a string, and the UR5 gripper opens automatically above the correct bin.

Figures

From Figure 2: The current policy collects experience on the robot into a replay buffer, the world model trains on re... — From Figure 2: The current policy collects experience on the robot into a replay buffer, the world model trains on replayed sequences, and an actor critic learns from imagined rollouts in the latent space of the world model. Data collection and learning run in parallel.

From Figure 3 (left): World model learning. The encoder fuses the sensor inputs into stochastic representations, a re... — From Figure 3 (left): World model learning. The encoder fuses the sensor inputs into stochastic representations, a recurrent state predicts their sequence, and a decoder reconstructs the inputs.

From Figure 3 (right): Behavior learning. Actions, rewards and values are predicted along rollouts in the latent spac... — From Figure 3 (right): Behavior learning. Actions, rewards and values are predicted along rollouts in the latent space of the world model, without decoding images.

From Figure 8: After being knocked over with a pole, the walking A1 adapts within 10 minutes to withstand pushes or q... — From Figure 8: After being knocked over with a pole, the walking A1 adapts within 10 minutes to withstand pushes or quickly roll back onto its feet.

Abstract

To solve tasks in complex environments, robots need to learn from experience. Deep reinforcement learning is a common approach to robot learning but requires a large amount of trial and error to learn, limiting its deployment in the physical world. As a consequence, many advances in robot learning rely on simulators. On the other hand, learning inside of simulators fails to capture the complexity of the real world, is prone to simulator inaccuracies, and the resulting behaviors do not adapt to changes in the world. The Dreamer algorithm has recently shown great promise for learning from small amounts of interaction by planning within a learned world model, outperforming pure reinforcement learning in video games. Learning a world model to predict the outcomes of potential actions enables planning in imagination, reducing the amount of trial and error needed in the real environment. However, it is unknown whether Dreamer can facilitate faster learning on physical robots. In this paper, we apply Dreamer to 4 robots to learn online and directly in the real world, without simulators. Dreamer trains a quadruped robot to roll off its back, stand up, and walk from scratch and without resets in only 1 hour. We then push the robot and find that Dreamer adapts within 10 minutes to withstand perturbations or quickly roll over and stand back up. On two different robotic arms, Dreamer learns to pick and place multiple objects directly from camera images and sparse rewards, approaching human performance. On a wheeled robot, Dreamer learns to navigate to a goal position purely from camera images, automatically resolving ambiguity about the robot orientation. Using the same hyperparameters across all experiments, we find that Dreamer is capable of online learning in the real world, establishing a strong baseline. We release our infrastructure for future applications of world models to robot learning.

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