Reinforcement learning has the potential to automate the acquisition of behavior in complex settings, but in order for it to be successfully deployed, a number of practical challenges must be addressed. First, in real world settings, when an agent attempts a tasks and fails, the environment must somehow reset so that the agent can attempt the task again. While easy in simulation, this could require considerable human effort in the real world, especially if the number of trials is very large. Second, real world learning is often limited by challenges in exploration, as complex, temporally extended behavior is often times difficult to acquire with random exploration. In this work, we show how a single method can allow an agent to acquire skills with minimal supervision while removing the need for resets. We do this by exploiting the insight that the need to reset-skills. We propose a general-sum game formulation that naturally balances the objective of resetting and learning skills, and demonstrate that this approach improves performance on reset-free tasks, and additionally show that the skills we obtain can be used to significantly accelerate downstream learning.

Additional Feedback: Line-by-line comments: Line 129 - Seem to be missing a comma after the \ldots . Line 177 - One theoretical point that seems hidden or ignored in this work is what this expectation for J {forward} (and J {reset}) really means. Because of the iterative and continuous "reset, forward, reset, forward, ..." nature of the task, this expectation is being (implicitly) taken after some arbitrary number of iterations between resets and forward episodes. This is perhaps fine if the initial states converge to some non-degenerate stationary distribution but this ignores the, very real, possibility of there being inescapable terminal states. E.g. if the reset policy always eventually throws the robot into a hole then the stationary distribution will always have the robot in this hole and thus nothing can be learned.

We consider the problem of optimally composing a set of primitives to tackle control tasks. To address this problem, we introduce Neo-FREE: a control architecture inspired by the Thousand Brains Theory and Free Energy Principle from cognitive sciences. In accordance with the neocortical (Neo) processes postulated by the Thousand Brains Theory, Neo-FREE consists of functional units returning control primitives. These are linearly combined by a gating mechanism that minimizes the variational free energy (FREE). The problem of finding the optimal primitives' weights is then recast as a finite-horizon optimal control problem, which is convex even when the cost is not and the environment is nonlinear, stochastic, non-stationary. The results yield an algorithm for primitives composition and the effectiveness of Neo-FREE is illustrated via in-silico and hardware experiments on an application involving robot navigation in an environment with obstacles.

In most contact-rich manipulation tasks, humans apply time-varying forces to the target object, compensating for inaccuracies in the vision-guided hand trajectory. However, current robot learning algorithms primarily focus on trajectory-based policy, with limited attention given to learning force-related skills. To address this limitation, we introduce ForceMimic, a force-centric robot learning system, providing a natural, force-aware and robot-free robotic demonstration collection system, along with a hybrid force-motion imitation learning algorithm for robust contact-rich manipulation. Using the proposed ForceCapture system, an operator can peel a zucchini in 5 minutes, while force-feedback teleoperation takes over 13 minutes and struggles with task completion. With the collected data, we propose HybridIL to train a force-centric imitation learning model, equipped with hybrid force-position control primitive to fit the predicted wrench-position parameters during robot execution. Experiments demonstrate that our approach enables the model to learn a more robust policy under the contact-rich task of vegetable peeling, increasing the success rates by 54.5% relatively compared to state-of-the-art pure-vision-based imitation learning. Hardware, code, data and more results would be open-sourced on the project website at https://forcemimic.github.io.

Reinforcement learning has the potential to automate the acquisition of behavior in complex settings, but in order for it to be successfully deployed, a number of practical challenges must be addressed. First, in real world settings, when an agent attempts a tasks and fails, the environment must somehow "reset" so that the agent can attempt the task again. While easy in simulation, this could require considerable human effort in the real world, especially if the number of trials is very large. Second, real world learning is often limited by challenges in exploration, as complex, temporally extended behavior is often times difficult to acquire with random exploration. In this work, we show how a single method can allow an agent to acquire skills with minimal supervision while removing the need for resets.

A common approach used to control robots is to program them with code to detect objects, sequencing commands to move actuators, and feedback loops to specify how the robot should perform a task. While these programs can be expressive, re-programming policies for each new task can be time consuming, and requires domain expertise. What if when given instructions from people, robots could autonomously write their own code to interact with the world? It turns out that the latest generation of language models, such as PaLM, are capable of complex reasoning and have also been trained on millions of lines of code. Given natural language instructions, current language models are highly proficient at writing not only generic code but, as we've discovered, code that can control robot actions as well.