DayDreamer: World Models for Physical Robot Learning

TL;DR

The paper demonstrates that modern world-model approaches can enable sample-efficient, real-world robot learning without simulators. By leveraging Dreamer with a latent-space world model (RSSM) and imagined rollouts, a single hyperparameter setting yields locomotion, manipulation, and navigation skills across four diverse robots. The approach combines offline-style world-model training with online actor-critic optimization, using lambda-returns and large latent rollouts, while maintaining a decoupled training process for stability. Results show rapid learning and adaptation: a quadruped walks within an hour and rebounds from perturbations in minutes; arms achieve near-human performance on visual pick-and-place, and a wheeled robot learns goal-directed navigation from pixels in a couple of hours. The work provides a strong, practical baseline for future real-world world-model robotics and offers open infrastructure for further exploration.

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.

Figures (10)

• Figure 1: To study the applicability of Dreamer for sample-efficient robot learning, we apply the algorithm to learn robot locomotion, manipulation, and navigation tasks from scratch in the real world on 4 robots, without simulators. The tasks evaluate a diverse range of challenges, including continuous and discrete actions, dense and sparse rewards, proprioceptive and camera inputs, as well as sensor fusion of multiple input modalities. Learning successfully using the same hyperparameters across all experiments, Dreamer establishes a strong baseline for real world robot learning.
• Figure 2: Dreamer follows a simple pipeline for online learning on robot hardware without simulators. The current learned policy collects experience on the robot. This experience is added to the replay buffer. The world model is trained on replayed off-policy sequences through supervised learning. An actor critic algorithm optimizes a neural network policy from imagined rollouts in the latent space of the world model. We parallelize data collection and neural network learning so learning steps can continue while the robot is moving and to enable low-latency action computation.
• Figure 3: Neural Network Training We leverage the Dreamer algorithm hafner2019dreamerhafner2020dreamerv2 for fast robot learning in real world. Dreamer consists of two neural network components. Left: The world model follows the structure of a deep Kalman filter that is trained on subsequences drawn from the replay buffer. The encoder fuses all sensory modalities into discrete codes. The decoder reconstructions the inputs from the codes, providing a rich learning signal and enabling human inspection of model predictions. A recurrent state-space model (RSSM) is trained to predict future codes given actions, without observing intermediate inputs. Right: The world model enables massively parallel policy optimization from imagined rollouts in the compact latent space using a large batch size, without having to reconstruct sensory inputs. Dreamer trains a policy network and value network from the imagined rollouts and a learned reward function.
• Figure 4: A1 Quadruped Walking Starting from lying on its back with the feet in the air, Dreamer learns to roll over, stand up, and walk in 1 hour of real world training time, without simulators or resets. In contrast, SAC only learns to roll over but neither to stand up nor to walk. For SAC, we also had to help the robot out of a dead-locked leg configuration during training. On the right we show training curves for both SAC and Dreamer. The maximum reward is 14. The filled circles indicate times where the robot fell on its back, requiring the learning of a robust strategy for getting back up. After 1 hour of training, we start pushing the robot and find that it adapts its behavior within 10 minutes to withstand light pushes and quickly roll back on its feet for hard pushes. The graph shows a single training run with the shaded area indicating one standard deviation within each time bin.
• Figure 5: UR5 Multi Object Visual Pick and Place This task requires learning to locate three ball objects from third-person camera images, grasp them, and move them into the other bin. The arm is free to move within and above the bins and sparse rewards are given for grasping a ball and for dropping it in the opposite bin. The environment requires the world model to learn multi-object dynamics in the real world and the sparse reward structure poses a challenge for policy optimization. Dreamer overcomes the challenges of visual localization and sparse rewards on this task, learning a successful strategy within a few hours of autonomous operation.