Despite great strides in language-guided manipulation, existing work has been constrained to table-top settings. Table-tops allow for perfect and consistent camera angles, properties are that do not hold in mobile manipulation. Task plans that involve moving around the environment must be robust to egocentric views and changes in the plane and angle of grasp. A further challenge is ensuring this is all true while still being able to learn skills efficiently from limited data. We propose Spatial-Language Attention Policies (SLAP) as a solution. SLAP uses three-dimensional tokens as the input representation to train a single multi-task, language-conditioned action prediction policy. Our method shows an 80% success rate in the real world across eight tasks with a single model, and a 47.5% success rate when unseen clutter and unseen object configurations are introduced, even with only a handful of examples per task. This represents an improvement of 30% over prior work (20% given unseen distractors and configurations). We see a 4x improvement over baseline in mobile manipulation setting. In addition, we show how SLAPs robustness allows us to execute Task Plans from open-vocabulary instructions using a large language model for multi-step mobile manipulation. For videos, see the website: https://robotslap.github.io

Therefore, we split the action prediction problem into two steps: (1) we predict a Most Relevant Point, or MRP, which tells us which region of the world the policy must attend to; and (2) we reactively predict actions which determine where the robot should move in relation to that MRP: for example, how to approach the handle of an oven and when to close the gripper to grasp it. These two operations are performed sequentially using a modified PointNet++ (Qi et al., 2017) model that we refer to as Attention-based PointNet (A-PointNet), shown in Figure 2. The MRP Predictor can then be agnostic to the position of the robot, instead focusing on the features of the object relevant to the overall task, while the Action Predictor can learn to focus on features relevant just to what the next action should be. For example, in Figure 7, the MRP Predictor learns to focus on the handle; the Action Predictor focuses on the angle of the oven door. Image Pre-Processing First we convert the RGB and depth images into a point cloud. We augment the point cloud of the current timestep with our context c, the point cloud from the beginning of the episode. This aids both in combating occlusion, as well as in disambiguating between similar observations that occur during different trajectories. To reduce compute, we crop the working area to 1m, and down-sample using grid pooling, with a resolution of 1cm for the current timestep and 2.5cm for the context. Specifically, we select a random point in each voxel, to reduce overfitting.

The ability for an agent to continuously learn new skills without catastrophically forgetting existing knowledge is of critical importance for the development of generally intelligent agents. Most methods devised to address this problem depend heavily on well-defined task boundaries, and thus depend on human supervision. Our task-agnostic method, Self-Activating Neural Ensembles (SANE), uses a modular architecture designed to avoid catastrophic forgetting without making any such assumptions. At the beginning of each trajectory, a module in the SANE ensemble is activated to determine the agent's next policy. During training, new modules are created as needed and only activated modules are updated to ensure that unused modules remain unchanged. This system enables our method to retain and leverage old skills, while growing and learning new ones. We demonstrate our approach on visually rich procedurally generated environments.

Progress in continual reinforcement learning has been limited due to several barriers to entry: missing code, high compute requirements, and a lack of suitable benchmarks. In this work, we present CORA, a platform for Continual Reinforcement Learning Agents that provides benchmarks, baselines, and metrics in a single code package. The benchmarks we provide are designed to evaluate different aspects of the continual RL challenge, such as catastrophic forgetting, plasticity, ability to generalize, and sample-efficient learning. Three of the benchmarks utilize video game environments (Atari, Procgen, NetHack). The fourth benchmark, CHORES, consists of four different task sequences in a visually realistic home simulator, drawn from a diverse set of task and scene parameters. To compare continual RL methods on these benchmarks, we prepare three metrics in CORA: continual evaluation, forgetting, and zero-shot forward transfer. Finally, CORA includes a set of performant, open-source baselines of existing algorithms for researchers to use and expand on. We release CORA and hope that the continual RL community can benefit from our contributions, to accelerate the development of new continual RL algorithms.