FORGE gives a humanoid one controller for walking, stance changes, wrist motion, and contact-aware manipulation. It estimates wrist interaction forces from joint motion and motor-torque estimates, so it can adapt during sustained contact without wrist-mounted force/torque sensors. Hardware tests showed under-8-centimeter wrist tracking error and under-10-centimeter foot-placement error, while a diffusion policy used it to execute lifting tasks across four payload conditions.
What did the researchers build?
FORGE is a whole-body controller for humanoid robots: one system that coordinates locomotion with arm and hand-task movements. Its command interface supports several levels of detail. An operator or higher-level planner can direct the robot toward a task area, adjust its stance, or provide a wrist trajectory. The controller aims to make those commands work together rather than treating walking and manipulation as separate jobs.
The system also addresses a problem that appears when a robot pushes, holds, or places an object: it needs to respond to contact forces, but wrist-mounted force/torque sensors add hardware and integration demands. FORGE estimates interaction force at each wrist using joint-state information and reconstructed motor-control torques, then uses that estimate to adjust wrist-position targets during sustained, relatively steady contact.
For whole-body motion, FORGE starts from a pretrained locomotion controller and keeps that lower-body expert fixed. A learned residual policy adds upper-body actions and corrective lower-body adjustments, allowing the robot to reach or manipulate while preserving a learned walking foundation. The controller was trained without relying on full-body human motion demonstrations.

What did FORGE achieve in tests?
Hardware tests covered walking, squatting and lifting, commanded foot placement, dynamic reaching, and recovery from disturbances. In a dynamic-reaching evaluation, the reported root-mean-square error (RMSE)—a measure of typical tracking error—was below 8 cm for three-dimensional wrist position and below 10 cm for planar foot position. Those figures indicate useful task-space tracking, though they do not by themselves establish performance across every robot or operating environment.
The researchers also tested a 30 N push at the wrist and a 50 cm body drag. Representative traces showed transient tracking errors followed by recovery, with bounded pelvis acceleration; the supplied results do not give a numerical recovery time or acceleration limit. In estimator comparisons, the hybrid force estimator reduced error against analytical-only and learned-only alternatives for the evaluated quasi-static contacts, but no exact error values are reported in the provided text.
| Evaluation | Reported result |
|---|---|
| Dynamic wrist tracking | 3D wrist-position RMSE below 8 cm |
| Dynamic foot placement | Planar foot-position RMSE below 10 cm |
| Perturbation tests | Recovery traces after a 30 N wrist push and 50 cm body drag |
| Downstream learning data | 320 demonstrations: 80 for each of four payload conditions |
| Hardware trials | 20 trials per payload condition |
A separate test trained one diffusion policy—a model that turns visual input and task state into action commands—on demonstrations from four payload conditions. FORGE served as the policy’s whole-body executor for squat-and-lift tasks. This shows that a learned planner could issue commands at multiple levels of detail to FORGE; the available results do not provide a task success rate or payload weights.

How does FORGE control a humanoid?
FORGE combines three pieces: a multi-level command interface, a contact-force estimation and adaptation loop, and a residual whole-body controller.
The force pathway starts with a model-based momentum observer. An observer is a calculation that estimates hidden physical quantities from measurements—in this case, external interaction forces. FORGE adds a learned correction trained with simulator-generated supervision. At deployment, that hybrid estimator uses arm-joint states and reconstructed low-level proportional-derivative (PD) motor torques. It does not require a wrist force/torque sensor. Its target is the low-frequency component of equivalent wrist force, making it most relevant to sustained, quasi-static contact rather than every rapid impact.
During contact, the estimated force feeds a virtual-target loop. The loop adjusts the wrist’s position target to account for the force at the wrist and, optionally, a commanded force the robot should apply. This gives the controller a way to adapt its motion while maintaining contact, instead of relying only on a fixed wrist trajectory.
For whole-body coordination, the pretrained locomotion expert supplies the baseline lower-body action. A reinforcement-learning policy adds residual actions: learned adjustments on top of that baseline. It controls the upper body and can make corrective lower-body changes. Keeping the locomotion expert frozen reduces the amount of risky exploration needed to learn manipulation, while still allowing the robot to shift its balance as a task demands. The researchers also report using command masking and asymmetric residual training to improve stability and task-space tracking compared with ablations—tests that remove or change parts of the method.

Why does FORGE matter for real-world robotics?
Many humanoid jobs require a robot to move and manipulate at the same time: approach a work area, set its feet, reach for an object, and adjust its motion once contact begins. A controller that coordinates these steps could simplify the handoff between navigation, manipulation, and contact handling. FORGE’s multi-granularity commands also create a potential bridge between low-level robot control and a higher-level planner that specifies a task without dictating every joint movement.
The sensorless wrist-force estimate could reduce dependence on specialized wrist hardware for the tested type of sustained contact. “Sensorless” here applies to wrist force/torque measurement—not to all sensing: the estimator still relies on joint states and reconstructed motor torques. Its contact adaptation may be useful for tasks such as steadying or lifting an object, but the tests do not establish performance for all materials, impacts, or industrial workflows.
The diffusion-policy evaluation is relevant because it treats FORGE as an executor beneath a visuomotor planner, rather than asking the planner to output every joint command. That separation could help teams develop task-level behavior while keeping balance and whole-body execution in a dedicated control layer. Engineers comparing platforms can browse humanoid robots on Robot Overflow or explore the broader used robots marketplace.
What remains unproven about FORGE?
The reported force-estimation advantage is limited to the evaluated quasi-static contacts. The supplied results do not include exact force-error values, so the size of the improvement over analytical-only or learned-only estimation cannot be quantified here. Fast impacts, changing contact geometry, and less predictable materials remain open tests.
The tracking figures are reported for hardware evaluations, but the available text does not provide full task success rates, recovery times, or the numerical pelvis-acceleration bounds. The diffusion-policy test uses four payload conditions and 320 demonstrations, yet the payload weights and cross-condition success rates are not included. Finally, a frozen locomotion expert can preserve stable gait behavior, but it may also constrain the motions available to the residual policy. Broader testing across platforms and tasks would clarify how much retraining or tuning deployment requires.
What should engineers know about FORGE?
Does FORGE require wrist force/torque sensors? No. Its wrist-force estimator uses arm joint states and reconstructed low-level PD torques, though the robot still needs those underlying measurements.
Can FORGE control walking and manipulation together? Yes. It combines navigation, stance, and wrist-pose commands in one whole-body controller, with learned corrections that can coordinate upper- and lower-body motion.
How accurate was wrist tracking in the reported test? Dynamic-reaching wrist-position RMSE was below 8 cm in three dimensions. That is a test result, not a guarantee for every task or platform.
Was FORGE tested with a learned visual planner? Yes. A diffusion policy used head-mounted RGB and task state to command FORGE across four payload conditions, based on 320 demonstrations.
What is the practical takeaway?
FORGE combines locomotion, reaching, and sustained-contact adaptation in a single humanoid control system, with hardware results showing sub-8 cm wrist and sub-10 cm foot tracking error in a dynamic-reaching test. Its strongest contribution is a practical control architecture for connecting task-level commands to whole-body motion, while force-estimation accuracy and broader deployment performance still need fuller reporting.
