flexiv_control

A unified, real-time control & safety layer for the Flexiv Rizon.
One action contract for teleoperation, MPC, reinforcement learning, and real-to-sim-to-real manipulation.

Landing page · Quick start · Architecture · Action contract · Safety

_{Community project — NOT affiliated with, endorsed by, or supported by Flexiv Robotics.}

flexiv_control is a thin, fast layer between a high-level decision maker — a policy, an MPC planner, a teleoperator — and a Flexiv Rizon arm. It takes a high-level action, turns it into a safe stream of setpoints at the control rate, and reports back what actually happened.

Every consumer speaks the same action contract, and every backend — a real Rizon via Flexiv RDK, a dependency-free fake backend, or a MuJoCo simulation — consumes the same safety-filtered setpoint stream. That is what makes one controller reusable across teleop, RL, MPC, and real2sim2real projects without forks. The design (a thin client over a real-time-capable loop, one action interface, sim + real behind a single backend switch) is the shape that Deoxys, Polymetis, frankapy, SERL/HIL-SERL, and LeRobot converged on (design rationale).

Features

• One action contract — CartesianChunk / CartesianDelta / JointChunk + GripperCommand, with absolute or relative-to-chunk-start poses, a predict-vs-execute horizon, and a quantified ExecutionResult.
• Safety is first-class — named, version-controlled SafetyProfiles and a microsecond per-tick SafetyFilter (workspace box, speed/jump caps, joint limits, contact-wrench stop, watchdog) that clips or rejects and reports.
• Sim ↔ real behind one switch — a MuJoCo backend (Jacobian IK, gravity compensation) with a faithful GN01 gripper, the fake backend for offline dev, and a real Rizon, all behind one config string.
• Built for learning & control — a Gymnasium env, a HIL-SERL intervention wrapper, a RecedingHorizonRunner (the VLA / MPC policy-server seam), and a SpaceMouse teleoperator, all on the same contract.
• Multi-user safe — an in-process lease and a host-wide lock so two processes can't fight over the arm.
• See what it WILL do — flexiv-control viz mirrors the robot live in any browser on the LAN and previews each chunk's intended motion (the true per-tick command path, time-colored, with gripper events, a workspace box, an animated ghost, and an optional Approve/Reject gate) before it executes — for safety and debugging. See docs/visualization.md.
• Optional, not required — a numpy-only cross-process server + RemoteRobot, a C++ 1 kHz real-time daemon, and a ROS 2 overlay. The core needs only numpy.

Installation

git clone https://github.com/ZihaoLu001/flexiv_control.git
cd flexiv_control
pip install -e .          # core: numpy only

pip install -e ".[rl]"        # + Gymnasium env
pip install -e ".[teleop]"    # + SpaceMouse
pip install -e ".[lerobot]"   # + LeRobot data/training
pip install -e ".[mujoco]"    # + MuJoCo simulation backend
pip install -e ".[viz]"       # + live browser visualization (viser)
pip install -e ".[flexiv]"    # + flexivrdk (real hardware; RDK v1.x)
pip install -e ".[dev]"       # + pytest, ruff

Quick start

No hardware needed — the fake backend is dependency-free:

from flexiv_control import Robot, RobotConfig, CartesianChunk

robot = Robot(RobotConfig(backend="fake"))     # or "mujoco" / "flexiv_rdk"
robot.connect()
robot.start_cartesian_impedance()

chunk = CartesianChunk.from_waypoint_array([[0.45, 0.0, 0.30, 1.0, 20],
                                            [0.50, 0.0, 0.25, 0.0, 20]])
result = robot.execute_cartesian_chunk(chunk)
print(result.success, result.path_tracking_error)

robot.disconnect()

flexiv-control demo            # offline FakeBackend demo, no hardware
python examples/01_fake_hello.py

Architecture

The action contract is the only thing crossing the boundary, so the network server, Gym env, and ROS overlay are all pure pass-through:

 Language policy   MPC planner   RL trainer   SpaceMouse
        \              |             |            /
         └──►  CartesianChunk / CartesianDelta / JointChunk  ◄──┘
                              │
            Robot facade → SafetyFilter (per-tick) → Interpolator → backend → Rizon

The Rizon runs its hard real-time impedance loop inside the robot, so the host only streams setpoints. That gives two tiers — a Python default (100–500 Hz, RDK non-real-time modes) and an optional C++ 1 kHz daemon (RDK real-time modes) — sharing the same contract. The landing page and docs/architecture.md lay this out in full.

Examples

Documentation

Safety

Before any motion on real hardware: keep an E-stop within reach and validate your SafetyProfile — especially the workspace box — on your own cell at low speed. See docs/safety.md.

Status

Alpha. The Python core, safety, action contract, server/client, Gymnasium env, teleop, and the MuJoCo simulation (including the GN01 gripper) are tested in CI (Python 3.8 / 3.10 / 3.12). The flexiv_rdk real-hardware backend is now HARDWARE-VALIDATED on a Rizon4s (serial Rizon4s-062626, RDK-Professional, flexivrdk 1.7) for the NRT pick-place path: enable → F/T-sensor zero → NRT_CARTESIAN_MOTION_FORCE (cartesian impedance) → NRT joint chunks → GN01 gripper → go_home_safe all run end-to-end (see CHANGELOG 0.1.5/0.1.4). Four RDK-1.7 API mismatches found there are fixed: gripper Enable(name), the Mode enum (1.7 has no RT_*), ZeroFTSensor-on-connect (force-mode prerequisite), and SendJointPosition's five list args. The C++ / ROS 2 paths remain unvalidated. RDK is version-matched to firmware — see docs/versions.md before use on a different setup.

Development

pip install -e ".[dev]"
ruff check src tests examples
pytest -q

Contributions welcome — see CONTRIBUTING.md.

Citation

If you use this in academic work, please cite it (see CITATION.cff).

License

Apache-2.0. The MuJoCo GN01 gripper model is derived from the official Flexiv ROS description (Apache-2.0); see NOTICE.