PHABS

Manipulation policies are trained on data that is blind to force. Human video carries no force signal, simulation invents it from designer-chosen contact parameters, and teleoperation usually withholds it from the demonstrator. For tasks where success is set by how hard you press, peeling a seal, opening a bag, handling something fragile, the demonstrations we train on are missing the variable that decides the outcome.

PHABS is my answer to that: a handheld bimanual teleoperation device built specifically to produce force-annotated demonstrations for imitation learning. It renders per-hand pinch force and, uniquely, the internal force between the two hands on a shared object, the signal that distinguishes crushing an object from merely supporting it, which no handheld bimanual device had rendered before.

This project is in active development. The paper is targeted at ICRA 2027 and the figures below show work in progress, not a finished system.

Left: the Anvil OpenArm dual-arm follower. Right: an early PHABS prototype held in both hands, the two pincher assemblies coupled through the gimbal rings and shared rails that carry the inter-hand squeeze channel (shown before electronics and actuators were installed).

Leading the project, I work from the data requirements backwards. What the learning pipeline needs determines the design: which channels are worth rendering, what has to be logged and at what rate, and which hardware compromises are acceptable because they don’t degrade the resulting dataset.

The software:

• Bilateral control loop on an Anvil OpenArm dual-arm robot: contact wrenches estimated from gravity-compensated joint effort (recursive Newton–Euler), mapped through the manipulator Jacobian, and rendered back to the operator as pinch and signed inter-hand force.
• Data pipeline capturing synchronized device-frame, robot-frame, and object-pose streams into a schema shared across collection, training, and deployment, so demonstrations are directly trainable rather than needing post-hoc reconstruction.
• Policy learning: ACT policies trained on matched haptic vs. no-haptic demonstrations across contact-rich bimanual tasks, an ablation designed to isolate what felt force at collection time is actually worth to the resulting policy.

The hardware, in one line: capstan-driven pinch actuators and a gimbal-motor squeeze stage, with force rendering verified against a reference load cell so the numbers in the dataset mean something.

Patent in preparation. First-author paper targeted at ICRA 2027.