Bimanual Redundancy Optimization

Paper: Task-Specific Manipulability Metrics for Redundancy Optimization in Cooperative Manipulation Authors: Debojit Das, Barat S., Harish J. Palanthandalam-Madapusi Status: Provisionally accepted, Industrial Robot: The International Journal of Robotics Research and Application Project website: https://debojit-d.github.io/Bimanual-Redundancy-Optimization/ Repository: https://github.com/Debojit-D/Bimanual-Redundancy-Optimization

Overview

Reference implementation for task-specific redundancy optimization in cooperative dual-arm manipulation: null-space redundancy optimization that preserves commanded object motion while shaping velocity, force, and directional-force manipulability. The manuscript evaluates this on both planar dual-arm hardware and a spatial dual-Franka MuJoCo simulation; this repository implements the spatial simulation study, config-driven end to end through a single CLI.

The older ROS 1, Gazebo, and MoveIt experiments that predate the MuJoCo implementation are retained under legacy/ for historical reference only and are not part of the active Python environment.

Video

Watch the hardware experiments and dual-Franka simulations

Implemented objectives

The optimization utilities implement four manuscript costs. "Directional force" is not one formula: the manuscript defines two distinct, non-equivalent objectives (Appendix A), and this repository implements both.

Velocity (Eq. 13, maximized):
    W_v = sqrt(det(A A.T))

Force (Eq. 14, maximized):
    W_f = sqrt(det((A A.T)^dagger))

Directional force, direct (Eq. 16, minimized):
    normalized Frobenius distance between (A A.T)^dagger and F

Directional force, indirect (Eq. 17, maximized):
    normalized Frobenius distance between A A.T and F

The direct formulation (--objective directional_force) compares in force-capability space and is the default/primary directional-force mode in this spatial study. The indirect formulation (--objective directional_force_indirect) compares in velocity-capability space and is additionally evaluated in the static and six-dimensional comparisons; it corresponds to the formulation used by the manuscript's planar hardware experiments (see Hardware implementation). Full equation-by-function detail is in docs/PAPER_CODE_MAP.md.

Quick start

git clone --recurse-submodules \
  https://github.com/Debojit-D/Bimanual-Redundancy-Optimization.git
cd Bimanual-Redundancy-Optimization

uv venv --python 3.12 .venv
source .venv/bin/activate

uv pip install -e .

Headless video recording additionally requires ffmpeg, and the interactive viewer requires a working OpenGL/display environment. If any step above doesn't work as shown (no python3.12, uv venv missing pip, no display), see docs/TROUBLESHOOTING.md.

5-minute example

Validate that the reference robot embodiment loads correctly:

bimanual-redopt validate-robot --robot dual_franka_panda

This prints the controlled joint count and the resolved qpos/DoF indices: a sanity check that the MuJoCo model, submodule, and environment are wired up correctly.

Then run a short, headless smoke pass of one paper campaign end to end (scene construction, controller, and CSV recording, at a tiny simulated duration, checking plumbing rather than publication statistics):

bimanual-redopt run --config configs/paper/static.toml --smoke

The command prints the output directory it wrote under outputs/paper_reproduction/.

Reproduce the paper

Every manuscript campaign is driven by a TOML config under configs/paper/ through the same CLI:

Paper config	Command	Output
static.toml	bimanual-redopt run --config configs/paper/static.toml	Static optimization campaign (Figures 10, 12)
six_d.toml	bimanual-redopt run --config configs/paper/six_d.toml	Representative 6D trajectories (Figure 11, part of Figure 14/Table I)
translational.toml	bimanual-redopt run --config configs/paper/translational.toml	Translational pick-and-place (Figure 13, part of Table I)
directional_direct_vs_indirect.toml	bimanual-redopt run --config configs/paper/directional_direct_vs_indirect.toml	Direct vs. indirect directional-force comparison (Figure 16)

Run every campaign in one call with bimanual-redopt reproduce-paper, or bimanual-redopt reproduce-paper --smoke for the short plumbing check above.

For the full manuscript-figure-level breakdown and output provenance, see docs/REPRODUCING_THE_PAPER.md. For per-script interactive/ad hoc experiment usage outside the config-driven pipeline, see docs/RUNNING_EXPERIMENTS.md.

Paper <-> Code

docs/PAPER_CODE_MAP.md is the authoritative equation index: every manuscript equation number maps to its implementing function, file, and test, for example:

Eq.	What	Function	File
8	Closed-loop redundancy update law	Equation8Controller.update	core/controller.py
13	Velocity manipulability	velocity_cost	core/objectives.py
16	Directional force (direct)	directional_force_direct_cost	core/objectives.py

Repository structure

src/bimanual_redundancy/   Active package: core/ (math), simulation/ (MuJoCo
                            backend), experiments/, plotting/
models/                     Robot and object MJCF models
configs/                    Paper reproduction configs, robot profiles
tests/                      Top-level pytest suite
docs/                       Architecture, equation map, reproduction guides
legacy/                     Archived ROS 1/Gazebo/MoveIt code (reference only)
mujoco_curobo_bridge/       Pinned external bridge submodule

See docs/ARCHITECTURE.md for the full package breakdown and dependency direction.

Hardware implementation

The manuscript also reports planar dual-arm hardware experiments. That hardware implementation is not distributed in this public repository, and the hardware results are not reproducible from this codebase; only the spatial MuJoCo simulation study is.

For research inquiries regarding access to the planar hardware implementation, please contact Barat S.: https://www.linkedin.com/in/baratsuresh2811/

Adding another robot

The framework supports additional cooperative robot embodiments through CooperativeSystemSpec, a minimal specification interface that separates the generic manipulability mathematics from any one robot's kinematics, actuation, or MuJoCo model. See docs/ADDING_A_ROBOT.md for the extension interface and checklist.

Contributions are welcome for additional robot embodiments, cooperative grasp configurations, new task-specific objectives, analytical/autodiff gradient implementations, and benchmarking/reproduction tooling. See CONTRIBUTING.md.

Acknowledgements

We thank Shail Jadav and Saniya Patwardhan for assistance with preliminary exploration of this work, and Samay Jain for assistance with collision handling in the spatial simulations.

Citation

If you use this code, please cite the associated manuscript. Machine-readable metadata is also available in CITATION.cff.

@article{das2026taskspecific,
  title   = {Task-Specific Manipulability Metrics for Redundancy Optimization in Cooperative Manipulation},
  author  = {Das, Debojit and S., Barat and Palanthandalam-Madapusi, Harish J.},
  journal = {Industrial Robot: The International Journal of Robotics Research and Application},
  year    = {2026},
  note    = {Provisionally accepted}
}

Development

After Quick start, install the test dependencies and run the suite from the repository root:

uv pip install -e ".[dev]"
pytest

See CONTRIBUTING.md for setup, branch naming, and guidelines for adding robots, objectives, or new mathematics.

Limitations

• The grasp is generated by physical fingertip contact; there is no weld, teleport, or hidden object attachment.
• The manuscript's planar hardware experiments are not distributed or reproducible from this repository (see Hardware implementation).
• Legacy ROS/Gazebo code requires a separate ROS 1 workspace and is not installed by the repository-local Python environment.

License

Original code in this repository is released under the Apache License 2.0. Third-party models, assets, dependencies, and submodules remain subject to their respective licenses. See THIRD_PARTY_NOTICES.md.