FORGE: Force-Aware Multi-Granularity Control for Efficient Humanoid Loco-Manipulation

The 30-second take

• FORGE runs locomotion and manipulation from sparse wrist poses, coarse motion commands, and an optional wrist-load target, not from a full-body demonstration.
• Sustained contact is estimated from proprioception and arm torques, so the robot does not need wrist force sensors to hold a push.
• Teleoperation labs that already drive humanoids can collect contact-rich behavior with a much thinner command interface.

What the paper actually did

The authors build FORGE, a force-aware controller whose commands run from coarse to fine: velocity navigation, planar foot placement, wrist-pose tracking, and optional force conditioning. Policy training does not use full-body human motion trajectories or joint-angle references. At deployment the policy consumes only sparse bilateral wrist poses relative to the pelvis, coarse motion commands, and optional desired wrist-load inputs. For sustained contact, a momentum observer is combined with a learned correction trained on simulator contact forces. On the robot, low-frequency wrist-equivalent interaction forces are estimated from proprioception and reconstructed low-level arm PD torques, without wrist force-torque sensors. A quasi-static virtual-target loop uses that estimate, and an optional desired applied force, to adapt the wrist target. A frozen locomotion prior and asymmetric residual actions stabilize the whole-body corrections that manipulation induces. The experiments cover unified-control stability, task-space tracking, quasi-static force estimation, squeezing a rigid box, and a downstream diffusion policy executing squat-and-lift across payload conditions.

What makes this disruptive

Contact-rich humanoid data is scarce because an operator is usually asked to specify the whole body at once, including balance. FORGE leaves walking to a frozen prior and lets a residual absorb the manipulation. The force signal comes from torques the arms already produce, not from extra sensors on the wrists. That is a thinner sensing and teaching stack for the same class of push, squeeze, and lift.

Why it matters (outside the lab)

A humanoid that can walk up to an object and keep a steady force on it is only useful if someone can demonstrate the task. Two wrist poses, a coarse motion, and an optional load are commands a teleoperator can actually send. The squat-and-lift test is the practical check: the same controller has to stay coordinated while the payload changes.

Limitations & open questions

The learned force correction is trained with simulator contact-force supervision, and the abstract does not claim the onboard estimate matches a wrist sensor on every contact. Squeezing is shown on a rigid box, and the downstream result is one squat-and-lift across payload conditions. Fast impacts, soft objects, and slips are outside what this abstract reports. A project site is named; a public failure table is not.