MEKAWIRE
ExpertarXiv

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

Liangjun Xing, Ke Fan, Kailin Li, Bin Wang, Xiaojie Niu, Qingwei Ben, Jiafei Cao, Wei Yao, Yebin Liu, Jingbo Wang, Weinan Zhang
In 30 seconds

FORGE is a preprint describing a humanoid controller that combines walking, foot placement, wrist targeting and sensorless contact feedback. Across 20 hardware trials per controller, wrist mean position error was 4.15 cm for FORGE versus 6.42 cm for SONIC, suggesting a more precise execution layer for teleoperation and learned manipulation.

Research question

Can one humanoid controller combine stable locomotion, precise task-space commands and sustained-contact adaptation without full-body human motion references or wrist force/torque sensors?

Problem

Manipulating objects while balancing requires arm movements, stance and contact forces to remain coordinated. Dense motion tracking does not directly provide precise task-space control, while rigid wrist-position tracking can squeeze objects during sustained contact. Estimating contact from joint signals is difficult because friction, actuation errors and whole-body motion also affect those signals.

Previous approach

Motion-reference policies rely on dense whole-body demonstrations. Modular locomotion and upper-body controllers, including HOMIE, can restrict coordination across the body. SoFTA separates slow locomotion from fast upper-body control. UniFP uses proprioceptive history for force estimation and position-target compensation, but sensorless estimates remain sensitive to model and actuator errors.

New approach

FORGE masks observations and rewards for inactive navigation, stance, wrist-pose and optional force commands. A two-layer Transformer processes six observable frames, alongside an adaptation encoder. A momentum observer uses arm states and reconstructed PD torques; a learned correction receives simulator force supervision during training. The resulting force estimate adjusts virtual wrist targets rather than commanding torque directly. A frozen locomotion policy supplies the gait, with bounded lower-body residuals and direct upper-body actions forming a 29-dimensional action. Control runs at 50 Hz. Training uses 2,048 parallel environments and up to 20,000 PPO updates per training stage, without full-body human motion references.

Results

In 20 hardware trials per controller, command-to-arrival latency was 22 ms for FORGE, 27 for HOMIE, 197 for TWIST2 and 252 for SONIC. Foot/wrist mean position errors were 5.42/4.15 cm for FORGE, 7.87/7.76 for TWIST2 and 7.35/6.42 for SONIC; HOMIE lacked the target interfaces for this comparison. In a separate simulated rigid-box test with identical wrist targets, squeezing forces were 8.52 N for FORGE, 15.48 for HOMIE, 17.16 for TWIST2 and 15.32 for SONIC. G1 force tests repeated each staged 0, 5, 10, 15 and 20 N load 10 times. On representative traces, x/y/z force RMSE was 0.81/1.97/0.87 N for the hybrid estimator, versus 2.83/5.15/3.02 N for observer-only and 2.43/2.23/2.88 N for learned-only estimation. Foam-box lifts showed less visible deformation with feedback, without a numerical deformation metric. A single Diffusion Policy trained on 320 demonstrations, 80 per payload condition, achieved 20/20, 20/20, 18/20 and 18/20 hardware transport successes for empty, 0.5 kg, 1 kg and 2 kg boxes, respectively: 100%, 100%, 90% and 90%.

Limitations

Hardware evidence is limited to G1 and the tested tasks. The estimator targets low-frequency, quasi-static translational wrist-equivalent forces, not contact moments or dynamic impedance control; it cannot separately identify contacts elsewhere on the arm. At held-out vertical loads of 35/40/50 N, hybrid RMSE rose to 8.28/9.66/12.35 N. Transport testing used only 20 trials per payload and no alternative executor baseline. Foam-box protection was assessed visually. Code availability and training runtime: Not reported in the paper.

Industry impact

Humanoid developers and warehouse or industrial handling teams could use this approach now as a research prototype for teleoperation, demonstration collection and execution beneath a visuomotor planner. Commercial use would require validation across more objects, contacts and operating conditions, including quantitative damage and reliability measurements. A commercial deployment timetable is not reported in the paper.

The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.