An Open-Source Robot-Training Backpack That Works Across Robot Bodies

A research team has open-sourced ModPack, a wearable rig for remote-controlling two-armed, mobile robots. The architectural move is decoupling the operator from any single robot body, so the same backpack can collect demonstration data across structurally different embodiments.

That decoupling is the contribution. Modern robot learning depends on humans physically demonstrating tasks, and the equipment that captures those demonstrations has historically been welded to one robot body. Swap the robot, rebuild the rig. ModPack's design treats the human side of that loop as a fixed unit and treats the robot side as a swappable module, which lowers the cost of collecting training data across robot families rather than within a single one.

The data side of robot learning is the live constraint. Manipulation policies improve with more high-quality demonstrations, but demonstrations are expensive: each one needs an operator, a session, and a rig configured for the specific robot. A rig that works for one arm pair on a fixed base does not, by default, work for a mobile base with a different arm stack. The standard workaround has been to build a new rig per embodiment, which multiplies the engineering work by the number of robots a lab wants to study.

ModPack pushes that engineering work into a shared wearable and a shared software interface, with the robot-specific hardware confined to swappable modules. The wearable is a 3D-printed backpack holding the onboard compute, batteries, communication hardware, and data storage an operator needs to drive a robot untethered. Inside, the design splits the load across five removable shelves running down the operator's spine, with excess material cut from structural panels to drop weight and let heat escape. The print spec is PLA at 15% gyroid infill with 0.2 mm layers, 2×0.4 mm walls, and tree supports; the project lists Bambu H2D and X1 Carbon as compatible printers.

The wearable is the platform; the modules are the point. ModPack exposes a shared interface that lets the same backpack host different capability modules depending on what the target robot needs. Operators can run the rig in joint-level teleoperation with haptic feedback, where the leader arms return resistance and contact forces to the operator's hands; in mobile manipulation mode, which adds base motion to the two-arm control loop; and in active perception, where the robot adjusts its own camera viewpoint during a task rather than relying on a fixed external camera.

The leader arms are six degrees of freedom and kinematically equivalent to the target arms, so the mapping between human and robot is one-to-one at the joint level rather than approximated through a learned correspondence. That detail does a lot of architectural work: a one-to-one joint map keeps the control loop legible, which is what makes the same interface usable across embodiments without per-robot retuning.

The authors report experiments on two structurally distinct robot platforms performing real-world mobile manipulation tasks and frame the system as a reusable framework for data collection and downstream policy learning. Policy learning, in this context, means training robot behavior from human-demonstrated trajectories, the same approach behind most of the recent progress in manipulation.

What ships is unusually complete for a robotics preprint. The project site releases the full hardware design and software stack, alongside a bill of materials with unit costs, quantities, and vendor links, plus step-by-step assembly documentation for the backpack and arms. The BOM is the kind of artifact that turns a paper into something another lab can rebuild over a weekend, which matters because the bottleneck ModPack is attacking is the rebuild step, not the demonstration step.

The caveats are real and worth stating in the open. The cross-embodiment and plug-and-play claims are author-asserted in a preprint, not a peer-reviewed result. The platform count is two, which is enough to show the interface works, not enough to show it generalizes. The published BOM prices and vendor links are point-in-time on the project's own site, not a market baseline. And teleoperation rigs themselves are not new, including for bimanual mobile work; the genuinely new piece is the wearable modular integration, not the underlying capability.

The next data point worth watching is whether another lab rebuilds the rig from the published hardware files and gets comparable cross-embodiment results on a third platform.