DEXTEROUS MANIPULATION

The world of dexterous hands.

Compare hardware. Discover applications. Share what works. An open reference and community for dexterous manipulation.

TriFinger Simulation

open-dynamic-robot-initiative

Official simulation environments and models for the TriFinger robot family.

What you can explore

Explore three-finger object manipulation and the documented connection between simulated and physical robot interfaces.

Before you use it

TriFinger is not an anthropomorphic hand. Match the robot variant, observation format and control conventions before transferring experiments.

Original contributors

Open Dynamic Robot Initiative contributors

License / access: BSD 3-Clause · check third-party assets separately

Original code & documentation

Research paper

FROM REFERENCE TO FIRST EXPERIMENT

TriFinger simulation: run a nine-joint check before training

INPUTModel-specific initial positions and position targets
OUTPUTSimulated joint observations; no learned policy
Environment
Python 3.10 / PyBullet
Embodiment
TriFingerPro in the example; three fingers, nine joints

What you need

Use a dedicated Python 3.10 environment, the version targeted by the installation documentation. The reviewed source also constrains NumPy below 2. Record the installed package versions and source commit before comparing results.

First useful check

Run a finite headless hold-position check. Add a GUI, objects and a policy only after the basic observation path works.

A small first experiment

  1. Create a separate environment

    Install the released package in a fresh environment. Keep the package manifest; a later release can differ from the source reviewed here. The source checkout is needed for repository demos and tests.

    python3.10 -m venv .venv-trifinger
    source .venv-trifinger/bin/activate
    python -m pip install --upgrade pip
    python -m pip install trifinger-simulation
    python -m pip check
    python -m pip freeze > trifinger-environment.txt
    Official installation guide →
  2. Choose the model before choosing the vector

    The introductory documentation uses FingerPro with three joints. TriFingerPro needs nine. Read the model-specific initial configuration instead of repeating a pose from a different finger type; the joint-direction conventions differ between models.

    Model names and initial poses →
  3. Run a finite headless simulation

    Save this original example as check_trifinger.py and run it in that environment. It issues simulated position commands for 200 steps and rejects malformed observations. Position targets pass through an internal PD controller; a successful run is not evidence of grasp stability.

    import numpy as np
    from trifinger_simulation import SimFinger, finger_types_data
    
    model = "trifingerpro"
    q = list(finger_types_data.get_initial_joint_positions(model))
    if len(q) != 9:
        raise ValueError("This example expects the nine-joint TriFingerPro model")
    sim = SimFinger(finger_type=model, enable_visualization=False)
    sim.reset_finger_positions_and_velocities(q)
    for _ in range(200):
        t = sim.append_desired_action(sim.Action(position=q))
        observed = np.asarray(sim.get_observation(t).position)
        if observed.shape != (9,) or not np.isfinite(observed).all():
            raise ValueError("Invalid simulated observation")
    print("Completed 200 simulation steps; last time index:", t)
    print("Last joint positions:", observed)
    print("Interface check only; no grasp score or hardware test")
    Simulation action and observation implementation →
  4. Understand what the returned time index means

    append_desired_action advances the simulation and returns an index. get_observation(t) is the state before that indexed action takes effect. In simulation, t + 1 can read the post-action state, but this shortcut does not transfer to the real-robot API. Keep action, index and observation together in logs.

    Time-index semantics →
  5. Add the platform and camera demo separately

    After cloning the upstream repository and installing that checkout, inspect demos/demo_trifinger_platform.py and its --help options. This broader demo adds the platform/object/camera path. GUI and rendering problems should be isolated from the headless physics check.

    Platform demo source →
  6. Turn the run into a reproducible experiment

    Record package versions, model, initial pose, action type, timestep, seed and object configuration. For a grasping result, add a defined success criterion, trial count and failure log; keep simulation results separate from physical trials.

    Plan a reproducible policy experiment →

What to record

  • The run finishes, reports nine finite joint positions and saves the environment manifest.
  • The model name, action units and observation timing are recorded before any policy comparison.
Common problems and what to check
Three values for a nine-joint model
Use the initial pose for trifingerpro; do not reuse the three-value FingerPro example.
Import or dependency conflicts
Use the isolated environment and pip check. The reviewed source requires NumPy <2; do not upgrade a working control environment to satisfy this tutorial.
No GUI on a remote machine
Keep visualization disabled for the first check. A headless physics run does not verify OpenGL or camera rendering.
A successful script is called a successful grasp
This example only holds an initial pose. It includes no object, task policy or success-rate evaluation.

Where it stops

TriFinger is a three-finger research platform, not an anthropomorphic five-finger hand. This example checks the simulation interface; it does not reproduce a benchmark or validate real-robot control.

Public documentation and source reviewed; no hardware or benchmark reproduction · reviewed 2026-10-24. Record the exact source revision when reproducing it.

Keep exploring

Gymnasium Robotics

MuJoCo robotics environments with Gymnasium interfaces, including dexterous hand tasks.

Faive Gym

Isaac Gym environments and training tools for the Faive hand.

Editorial reference · checked 2026-10-24

Start a discussion

Published posts are public. First contributions are reviewed. Only share material you have permission to disclose.

Report a post