Parallel MPC cluster manager, based on a client-server architecture, for use with high-throughput vectorized simulation and more. Designed for data-hungry applications, like RL-Augmented MPC (see AugMPC).

Ships with many useful tools for MPC development, including an extensible GUI for real-time debug.

_{(robot visualization on the left comes from MPCViz)}

Why MPCHive

• Throughput: from hundreds up to thousands of parallel MPC controllers (on CPU) for RL data collection, batch benchmarking, MPC design and more.
• Determinism: careful shared-memory implementation and MPC synchronization ensure reliable data (see EigenIPC).
• Modularity: Clean separation between ControlClusterServer, ControlClusterClient (process manager), and controller implementations.
• Flexibility: drop in your own MPC formulations by subclassing RHController and ControlClusterClient; mix different clusters for different robots or formulations.
• Real/Sim parity: same shared-memory schema for simulation, hardware-in-the-loop, and real deployments.

Architecture

• ControlClusterServer (mpc_hive.cluster_server.control_cluster_server). It's the interface with the simulator/real robot and creates shared-memory data servers for RobotState, RhcCmds, RhcPred, RhcPredDelta, RhcRefs, RhcStatus, and so on. The whole cluster runs at a fixed cluster_dt, static across all MPCs. Multiple clusters can be used to circumvent this limitation.
• ControlClusterClient (mpc_hive.cluster_client.control_cluster_client): spawns a pool of processes (one for each MPC) and manages their lifetime. MPCs register to the cluster and have assigned a unique id, which is then used to properly access data when reading robot states and writing solutions. The ControlClusterClient._generate_controller() method should be overriden to return a copy of your specific controller.
• Controllers: All MPC implementation should inherit from the mpc_hive.controllers.rhc.RHController base class. When triggered remotely by ControlClusterServer.trigger_solution(), controllers read the state of the robot, run the solver algorithm, write the solution, and update profiling/debug data, all using their unique registration index within the cluster.
• Shared data utilities (mpc_hive.utilities.shared_data.*): EigenIPC-backed tensors/dicts abstractions for all robot/cluster data; GPU mirrors based on Torch are optionally made available for interfacing with GPU simulators.
• Utilities: debugger GUI, keyboard/joy teleop, etc..

Installation

The preferred way to install MPCHive is through IBRIDO's container, which ships with all necessary dependencies. To setup the container, follow the instructions at ibrido-containers.

Quick start

The easiest way to get started is to run one of the examples provided by the AugMPC project, following the instructions available at ibrido-containers, which already demonstrate the integration with vectorized simulators and implementations of specific MPC controllers.