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

November 2026

TL;DR One RL policy for a G1 humanoid takes velocity, foothold, wrist-pose and wrist-force commands, and estimates wrist contact forces from arm joint torques without force sensors. It cuts wrist tracking error to 4.15 cm (vs 6.42-7.76 cm for TWIST2 and SONIC), and a diffusion policy driving it carries boxes of up to 2 kg in 18-20 of 20 trials.

What they do

FORGE is a single whole-body RL policy for a 29-DoF G1 humanoid that accepts base velocity, planar footholds, bilateral wrist poses and an optional desired wrist force, alone or combined. It is trained without full-body human motion references.

How it works

Masked command groups. In training, inactive command groups have their observations nulled and their rewards removed, so one policy serves every combination.
Sensorless wrist force. A momentum observer on the arms, fed by PD torques reconstructed from joint targets and states, estimates wrist force; a learned head, supervised with simulator contact forces, corrects its error.
Virtual-target contact loop. The gap between desired force and estimated reaction shifts the wrist position target (200 N/m).
Asymmetric residual. A pretrained locomotion policy is frozen; a residual policy drives the upper body and adds small, clipped leg corrections.

Results

Against HOMIE, TWIST2 and SONIC over 20 trials each, FORGE has the lowest latency (22 ms vs 27, 197 and 252 ms), foot error (5.42 cm vs 7.35-7.87 cm) and wrist error (4.15 cm vs 6.42-7.76 cm). Squeezing a rigid box in simulation, it applies 8.52 N vs 15.3-17.2 N.
Removing the masking raises wrist error from 0.079 to 0.115 m and shortens valid rollouts from 19.5 to 15.7 s.
On hardware with gauge-measured loads, the hybrid estimator’s per-axis errors are 0.81-1.97 N, vs 2.83-5.15 N for the observer alone and 2.23-2.88 N for a learned-only estimator.
A diffusion policy trained on 320 demonstrations commands FORGE through squat-and-lift box transport: 20/20 empty, 20/20 at 0.5 kg, 18/20 at 1 kg and 18/20 at 2 kg.

Limitations

The estimator recovers only the low-frequency translational force at each wrist, not a full wrench, and does not separate contacts on other links; at held-out 35/40/50 N loads its RMSE rises to 8.28/9.66/12.35 N. The force loop is quasi-static, not dynamic impedance control. The controller comparison for contact force is in simulation, and the downstream policy is tested on one task.

Figures

From Figure 2: The FORGE pipeline: one policy for navigation, stance, wrist-pose and force commands, a hybrid force e... — From Figure 2: The FORGE pipeline: one policy for navigation, stance, wrist-pose and force commands, a hybrid force estimator with a virtual-target contact loop, and residual actions on a frozen locomotion prior.

From Figure 3: The actor. A Transformer reads six masked frames of commands and proprioception; an adaptation branch ... — From Figure 3: The actor. A Transformer reads six masked frames of commands and proprioception; an adaptation branch outputs a 6-D correction to the momentum-observer force estimate. Privileged simulator quantities supply training targets only.

From Figure 8: Real-robot x/y/z wrist-force estimates under staged 0, 5, 10, 15 and 20 N loads: the hybrid estimator,... — From Figure 8: Real-robot x/y/z wrist-force estimates under staged 0, 5, 10, 15 and 20 N loads: the hybrid estimator, observer-only and data-driven variants against a force-gauge reference.

From Figure 9: Lifting the same foam box with the force-conditioned virtual-target update (left) and with plain posit... — From Figure 9: Lifting the same foam box with the force-conditioned virtual-target update (left) and with plain position tracking (right), which deforms the box more.

Abstract

Humanoid loco-manipulation requires coordinated locomotion, balance regulation, and manipulation under frequent contacts, making teleoperated demonstration collection difficult: a practical controller must simultaneously provide precision, stability, and contact awareness. We propose FORGE, a force-aware multi-granularity control system with a coarse-to-fine command space spanning velocity navigation, planar foot placement, wrist-pose tracking, and optional force-conditioning inputs. Policy training requires no full-body human motion trajectories or joint-angle references; deployment consumes only sparse bilateral wrist poses relative to the pelvis, coarse motion commands, and optional desired wrist-load inputs. For sustained contact, FORGE combines a momentum observer with a learned correction trained using simulator contact-force supervision. At deployment, it estimates low-frequency, wrist-equivalent interaction forces from proprioception and reconstructed low-level arm PD torques, without wrist force/torque sensors. A quasi-static virtual-target loop uses this estimate and an optional desired applied-force input to adapt the wrist target. A frozen locomotion prior and asymmetric residual actions stabilize manipulation-induced whole-body corrections. Experiments evaluate unified-control stability, task-space tracking, quasi-static wrist-equivalent force estimation, squeezing forces during sustained contact with a rigid box, and downstream diffusion-policy execution of a squat-and-lift task across payload conditions. Website: https://bb0928.github.io/forge/.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong