Robotics Environment Authoring Framework (REAF)

The Robotics Environment Authoring Framework (REAF) simplifies creating environments that adhere to the GDM Robotics Environment interface.

How to install

reaf can be installed from PyPI using pip:

pip install reaf

Directory Structure

Currently, the directory structure is designed as standalone subdirectories with a well-defined dependency graph.

reaf
├── core: Core libraries and interfaces for REAF.
├── testing: General tooling for testing REAF interfaces and environments.
├── common: Libraries with shared functionality across setups and platforms.

Design

REAF is a framework designed to simplify the creation of robotics environments. It adopts a layered architecture to promote modularity and reusability. The core components of a REAF environment are:

1. Environment: The top-level interface for interacting with the environment, conforming to the GDM Robotics Environment interface. It handles stepping, resetting, and action/observation specs.
2. Task Layer: Responsible for defining the task itself, including reward calculation, termination conditions, features generation, and commands processing.
3. Device Layer: Interfaces with the physical or simulated robotic setup, managing commands to actuators and retrieving measurements from sensors.
4. Adapters: Bridge the gap between the abstract GDMR interfaces and the specific requirements of the Task Layer. These adapters translate agent actions into Task Layer commands and Task Layer features into agent observations.
5. Reset and End of Episode Handlers: Support customized behavior during environment resets and episode termination.

Environment

The Environment class serves as the primary interface for interacting with the robotic environment. It coordinates the interactions between the Task Layer and Device Layer, manages the environment's state, and handles stepping through the environment. Key functionalities include:

• reset_with_options(): Resets the environment to a new initial state based on the provided options. This involves resetting the Device Layer, computing initial features and observations.
• step(): Advances the environment by one step. This method takes an agent action, processes it into commands, steps the Device Layer, computes features, reward, discount, termination conditions, and new observations, and returns a TimeStep object containing this information.
• action_spec(): Returns the specification for valid agent actions. This is determined by the ActionSpaceAdapter.
• timestep_spec(): Returns the specification for the TimeStep objects returned by step() and reset().
• Logging: Facilitates adding and removing loggers to monitor internal operations.

Task Layer

The Task Layer defines the logic and rules governing the robotic task. It comprises several core components:

• FeaturesProducer: Generates additional features based on existing features and measurements from the Device Layer. Each producer has a produced_features_spec() defining the features it generates and required_features_keys() indicating the features it depends on.
• CommandsProcessor: Modifies commands before they are sent to the Device Layer. Processors can transform, filter, or augment commands. The consumed_commands_spec() describes the commands accepted by the processor, and produced_commands_keys() defines the output commands.
• RewardProvider: Calculates the reward signal based on the current features. It exposes a reward_spec() defining the structure of the reward.
• TerminationChecker: Determines whether the episode should terminate based on features and returns a TerminationResult indicating the type of termination.
• DiscountProvider: Computes the discount factor based on features and the termination state.
• FeaturesObserver: Passive components that observe features without modifying them. This is useful for logging or analysis.
• Logger: Records measurements, features, and commands during environment interactions. Methods like record_measurements(), record_features(), and record_commands_processing() are called at specific points in the environment's lifecycle.

The Task Layer also provides methods to:

• compute_all_features(): Computes all features based on measurements from the Device Layer and the output of FeaturesProducers.
• compute_final_commands(): Processes the policy's commands using the CommandsProcessors and outputs the Device Layer command.
• compute_reward(): Calculates the reward using the RewardProvider.
• check_for_termination(): Checks termination conditions using TerminationCheckers.
• compute_discount(): Computes the discount using the DiscountProvider.
• validate_spec(): Verifies the consistency of the specs across the Task Layer and Device Layer.

Device Layer

The Device Layer serves as the bridge between the REAF environment and the robotic hardware. It's responsible for sending commands to the robot and receiving measurements from sensors. The Device Layer is built around:

• Device: Represents a single hardware component (e.g., robot arm, camera). It provides methods like set_commands() and get_measurements() for interacting with the hardware.
• DeviceCoordinator: Manages a collection of Device objects, coordinating their actions and data exchange. It provides lifecycle management through start() and stop() methods and synchronization points through before_set_commands() and before_get_measurements().

The Device Layer's key functions are:

• begin_stepping(): Initializes the Device Layer and returns the initial measurements.
• step(): Sends commands to the devices, retrieves new measurements, and returns them.
• end_stepping(): Performs cleanup operations at the end of an episode.
• commands_spec(): Returns the specification for valid commands. The user is expected to pass the full command dictionary.
• measurements_spec(): Returns the specification for the measurements returned by get_measurements().

Adapters

REAF utilizes adapters to translate between the generic agent interface and the specific format required by the Task Layer.

• ActionSpaceAdapter: Converts the agent's actions into a commands dictionary understood by the Task Layer.
• ObservationSpaceAdapter: Transforms the features generated by the Task Layer into observations suitable for the agent.

Reset and End of Episode Handler

• EnvironmentReset: Defines the reset behavior of the environment, including a do_reset() method and a default reset configuration. The reset is a complex step and very hard to predict how this will be written. The intention is to leave complete control to the user so they can pass whatever object is needed to the reset without any restrictions. The user can use this to check that everything is working properly, reset the state of any features producers, etc...
• EndOfEpisodeHandler: Provides a callback function, on_end_of_episode_stepping(), that is invoked at the end of each episode after the last step. This is useful for logging, cleanup, or custom logic that needs to be executed when an episode ends.

Licence and Disclaimer

Copyright 2026 Google LLC

All software is licensed under the Apache License, Version 2.0 (Apache 2.0); you may not use this file except in compliance with the Apache 2.0 license. You may obtain a copy of the Apache 2.0 license at: https://www.apache.org/licenses/LICENSE-2.0

All other materials are licensed under the Creative Commons Attribution 4.0 International License (CC-BY). You may obtain a copy of the CC-BY license at: https://creativecommons.org/licenses/by/4.0/legalcode

Unless required by applicable law or agreed to in writing, all software and materials distributed here under the Apache 2.0 or CC-BY licenses are distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the licenses for the specific language governing permissions and limitations under those licenses.

This is not an official Google product.