On November 12, 2026, a research team from the Shanghai AI Laboratory, Tsinghua University, Zhejiang University, Shanghai Jiao Tong University, and The Chinese University of Hong Kong released FORGE, a force-aware multi-granularity control framework for humanoid robots that resolves long-standing conflicts between locomotion balance, task-space positioning precision, and sustained contact regulation.
Humanoid loco-manipulation has historically posed a severe trade-off during teleoperation and imitation learning. Operating humanoids requires coordinating leg locomotion and body balance with dual-arm manipulation under continuous, dynamic physical interactions. Existing whole-body teleoperation stacks have typically forced researchers to choose between monolithic full-body motion capture tracking and decoupled architectures. Dense motion tracking policies often suffer from cumulative joint error and lack direct interfaces for contact force adaptation, while decoupled locomotion and manipulation stacks struggle with whole-body dynamic coordination during heavy manipulation tasks.
Multi-Granularity Command Architecture
FORGE (Force-Aware Multi-Granularity Control for Efficient Humanoid Loco-Manipulation), detailed in a preprint hosted on arXiv:2611.17142, introduces a unified command interface across four hierarchical operational levels: base navigation velocity, planar foothold placement, bilateral six-degree-of-freedom wrist poses, and optional target interaction forces. Crucially, the policy requires no full-body human joint trajectories or dense human motion reference data during training, drastically reducing the demonstration burden for embodied AI research laboratories.
To prevent conflicts between simultaneously active and inactive control channels, the system introduces dynamic command masking. When higher-level commands such as precise foot placement or specific applied loads are idle, their corresponding observation channels and reward functions are masked. This design eliminates gradient interference and stabilizes policy learning during transitions between fast omnidirectional walking and stationary high-precision manipulation.
Sensorless Force Estimation and Contact Adaptation
A major bottleneck in physical humanoid interaction is the fragility and cost of wrist six-axis force and torque sensors. FORGE bypasses hardware load cells entirely by deploying a physics-informed hybrid force estimator that operates on standard joint proprioception and reconstructed actuator proportional-derivative torques.
The estimator pairs a model-based generalized momentum observer with a learned neural compensation network. The momentum observer tracks low-frequency joint-space disturbance residuals mapped to the end effectors through the upper-body manipulator Jacobian. The neural network, trained with privileged ground-truth contact forces in simulation environments similar to robosuite and Isaac Gym, predicts unmodeled actuator friction, link compliance, and higher-order dynamic effects. A quasi-static virtual-target loop continuously shifts commanded wrist reference positions based on estimated contact forces, preventing excessive contact forces that could deform manipulated objects or destabilize the bipedal base.
Benchmarking Performance on 29-DoF Hardware
The research team evaluated FORGE on a 29-degree-of-freedom Unitree G1 humanoid robot across simulation environments and physical hardware trials. In comparison against established humanoid loco-manipulation baselines—including HOMIE, TWIST2, and SONIC—FORGE demonstrated superior operational latency, tracking accuracy, and contact force regulation during bilateral box interaction tests.
| Controller Method | Hardware Latency (ms) | Planar Foot MPE (cm) | 3D Wrist MPE (cm) | Box Squeezing Force (N) |
|---|---|---|---|---|
| HOMIE | 27 | N/A | N/A | 15.48 |
| TWIST2 | 197 | 7.87 | 7.76 | 17.16 |
| SONIC | 252 | 7.35 | 6.42 | 15.32 |
| FORGE (Ours) | 22 | 5.42 | 4.15 | 8.52 |
As detailed in the empirical evaluation, FORGE achieved an end-to-end command latency of 22 ms on physical hardware, representing an 88% reduction in latency compared to TWIST2 (197 ms) and an 91% reduction compared to SONIC (252 ms). In sustained bilateral contact benchmarks against a rigid box, FORGE maintained a mean contact force of 8.52 N, cutting the excessive squeezing forces observed in SONIC (15.32 N) and TWIST2 (17.16 N) by nearly half.
Implications for Demonstration Collection and Policy Scaling
Beyond low-level teleoperation stability, the framework provides an execution backbone for downstream visuomotor learning. In validation trials, the authors trained a 3D Diffusion Policy via imitation learning using multi-granularity command chunks generated by FORGE. Across a challenging squat, lift, and transport task evaluated over 80 hardware trials spanning payload variations from 0 kg to 2 kg, the unified policy achieved a 100% completion rate for empty and 0.5 kg payloads and a 90% success rate under 1 kg and 2 kg loads.
For university robotics facilities and research offices evaluating capital investments in robotic platforms, FORGE indicates that robust contact-rich manipulation does not mandate fragile, expensive multi-axis wrist force sensors. By open-sourcing policy checkpoints and teleoperation pipelines, the project establishes a reference model for conducting a reproducible AI experiment in full-body humanoid robotics without requiring custom sensorized end effectors.
Sources
- Xing, L., Fan, K., Li, K., Wang, B., Niu, X., Ben, Q., Cao, J., Yao, W., Liu, Y., Wang, J., and Zhang, W. (2026). FORGE: Force-Aware Multi-Granularity Control for Efficient Humanoid Loco-Manipulation. arXiv:2611.17142. https://arxiv.org/abs/2611.17142 (Checked 2026-11-15)
- FORGE Research Team. (2026). FORGE: Force-Aware Multi-Granularity Control Project Repository and Benchmark Datasets. https://bb0928.github.io/forge/ (Checked 2026-11-15)
