Ctrl-F: Force as an Interface for Humanoid Loco-Manipulation

Mengjie Zhao, Yuxin Chen, Yiyang Shao, Chuye Hong, Jiakang Chen, Koushil Sreenath, Masayoshi Tomizuka

November 2026

Raise-walk-kick on the G1, carrying a 1 kg box and kicking an 8 kg cart.

TL;DR Ctrl-F gives a humanoid motion tracker a 3-D force command for each hand or foot, trained without objects, plus a planner that sets those forces from the object's state. In simulation it keeps hand-task completion above 90% at 1.5x nominal object mass, where force-blind trackers drop to zero, and a real Unitree G1 pushes and kicks 20 kg boxes.

What they do

Humanoid motion trackers reproduce where the body should be, not how hard it should push. In the authors’ tests a force-blind tracker cannot lift 1.5-6 kg boxes because they slide off its palms. Ctrl-F adds force as an explicit command, a 3-D force for each designated contact body (the hands, a foot), on top of whole-body motion tracking. A high-level planner sets those forces online from the object’s state.

How it works

Inverse-force training. The low-level tracker is trained without objects. When a load mask marks a hand or foot as free, a force command is sampled and the opposite wrench is applied to that body, so the policy must exert the commanded force to keep tracking the reference.
Force planner. With the tracker frozen, a planner maps object state, reference motions and proprioception to force commands. It is trained with PPO plus a decaying DAgger-style imitation loss toward an analytical model. That model computes the needed object wrench from privileged pose and mass and splits it across the contact bodies.
Deployment. On a Unitree G1 with AprilTag object poses, the policy uses neither force sensing nor object mass.

Results

Simulation, five hand tasks: completion stays above 90% up to 1.5x nominal mass. At nominal mass the force-blind SONIC tracker completes 31% and Ctrl-F without its command 43%, and both reach zero by 1.5x.
Behavior cloning alone falls short: it completes 6% of box raises against 100% for the full recipe. In box lifting it commands 37 N, but only about 8 N reaches the box.
Hardware: commands of 10/20/30 N read 8.8/16.7/24.5 N on a crane scale. Ctrl-F drags a locked cart carrying 14 kg (and fails at 20 kg). It pushes and kicks a 20 kg box that SONIC does not move, and raises a 2 kg bin in 9 of 10 trials where SONIC fails. When weight is added mid-drag it keeps its pace, while a frozen force command slows down.

Limitations

Object state comes from AprilTag markers, so deployment depends on reliable marker detection. Each task family uses its own policy pair trained on a limited set of motions, and new clips need retraining. On the crane scale the delivered force fell 12-18% short of the command.

Videos

The hand pulls on a crane scale as the force command steps through 10, 20 and 30 N.

Hardware side-by-side with the force-blind SONIC tracker, starting with sweeping a 1 kg box.

Figures

From Figure 1: Ctrl-F on hardware, running the raise-walk-kick sequence with a 1 kg box in the hands and an 8 kg cart... — From Figure 1: Ctrl-F on hardware, running the raise-walk-kick sequence with a 1 kg box in the hands and an 8 kg cart to kick. The hands slide off a box they do not press into, so a tracker without a force command does not lift it at all.

From Figure 2: Stage 1 trains the motion-force tracker without objects: a force command is sampled and the opposite f... — From Figure 2: Stage 1 trains the motion-force tracker without objects: a force command is sampled and the opposite force is applied to the robot, so the tracker learns to follow its reference while exerting the commanded load. Stage 2 freezes the tracker and learns the force command from object and robot state, supervised by an analytical force model and refined by reinforcement learning.

From Figure 6: Left: task performance as object mass increases, averaged over the five hand tasks; 1.0x is each objec... — From Figure 6: Left: task performance as object mass increases, averaged over the five hand tasks; 1.0x is each object's nominal mass and the dashed line is 90% completion. Right: success at nominal mass, per task, for the hand policy.

From Figure 7: Hardware deployment on a Unitree G1 using onboard sensing and AprilTag object poses. Hand tasks: dragg... — From Figure 7: Hardware deployment on a Unitree G1 using onboard sensing and AprilTag object poses. Hand tasks: dragging the cart, pushing the box and raising the bin. Foot tasks: kicking the box and sweeping a 1 kg box. Composite: raise-walk-kick. The pushed and kicked boxes reach 20 kg, the heaviest load tested.

Abstract

Humanoid motion tracking has scaled to general policies that reproduce diverse human motion, but the resulting behavior is in free-space: the tracker knows where the body should be, not how it should interact with the environment. Other work makes the object part of the objective, yet abstracts that interaction to contact or geometry, which leaves the demanding cases under-explored. We argue that force is a natural command channel between motion imitation and object interaction, one that specifies not just where to make contact but how hard. We present Ctrl-F, a hierarchical motion-force framework that augments motion tracking with an explicit force command, a 3-D force at each designated contact body such as a hand or a foot. The hierarchical framework comprises: (1) A unified motion-force controller tracks the reference motion while applying the commanded force to the world. (2) A high-level force planner that issues a force command online based on object state and proprioception observations. In simulation and on hardware, Ctrl-F completes loco-manipulation that demands large, sustained interaction force and adjusts its command online as the load changes, both out of reach of a force-blind tracker. On hardware, Ctrl-F demonstrates hand-based dragging, pushing, and lifting, as well as foot-based kicking, successfully manipulating boxes weighing up to 20kg in pushing and kicking tasks where the force-blind baseline fails, while adapting online to additional loads introduced mid-task.

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