I-BFM

I-BFM lets a single humanoid policy carry, push, and kick boxes, using the LOGO method to serve both the contact at hand and the task goal further away.

When carrying a box, the robot must keep its body stable, hold the box, and move toward the target. Once the box deviates from its expected position, or the robot is disturbed on the way, the next contact has to be adjusted accordingly. Whether the task can still be finished tests an interaction policy more than completing a pre-arranged motion does.

Two intertwined difficulties sit here: carrying, pushing, and kicking require different ways of contacting the body with the object; and each step must serve the final goal. I-BFM first learns reusable body–object interaction capability, then steers the same policy with task rewards. Its LOGO method supplies both a nearby interaction direction and a slightly more distant goal direction, so the policy can keep advancing the task while adjusting contact.

In the MuJoCo carry-box evaluation on a Unitree G1, I-BFM reaches a 94.3% task success rate without disturbance; after the robot receives a downward external force during carrying, the success rate is 89.3%. This result comes from controlled disturbance tests in simulation, and the real-robot part has separate interaction demonstrations.

Why must interaction consider the body and the box at the same time?

This is not simply attaching an arm motion to a walking policy. When the robot holds the box, the body's centre of mass, the two-handed contact, and the box position change together; readjusting hand contact may also change the subsequent walking direction. Pushing and kicking use different contact locations and force application, and also require the policy to judge how to keep approaching the goal after the box moves.

Planning a trajectory and then tracking it is an effective technical route in humanoid interaction. I-BFM instead chooses to learn a shared closed-loop policy: every step decides the action from the current observation, so that walking, contact, and moving the box serve the same task goal. The research focus therefore falls on two things: how to let the policy cover different interactions, and how to serve both the contact at hand and subsequent task progress.

I-BFM: one policy covering multiple interactions

I-BFM follows the idea of behaviour foundation models: first learn reusable body control capability, then elicit the corresponding behaviour with task conditions. It also brings object motion and contact states into learning. Pretraining combines environment exploration through unsupervised reinforcement learning with unlabelled walking and interaction motion data. The model learns the body, object, and contact states that may appear after taking an action; the paper calls this forward–backward representation (FB).

When a new task arrives, researchers define a reward, for example whether the box is close to the goal or whether the robot stays stable. The model maps the reward to a latent instruction, and the same low-level policy then executes carrying, pushing, and kicking, without re-optimising a policy for each task.

LOGO: making the contact at hand work with the more distant goal

With a task direction in place, two time scales still have to be handled: how to contact the box next, and how to advance the task afterwards, may require different actions. To this end, the research proposes LOGO (Local–Goal Objective Geometry Operator): it extracts the immediately next "local goal" and the slightly more distant "task goal" from the same interaction trajectory, then supplies the direction and distance from the current state to both goals to the same policy. Put plainly, the policy knows both "how the hands and feet should move right now" and "where this segment of motion should advance toward".

LOGO's effect has direct ablation evidence. Under the same Carry Box setting, the full I-BFM reaches a 94.3% success rate with a final box error of 0.15 metres; removing LOGO and using an average of 8-step future features instead gives 51.0% and 0.33 metres. At the same time, joint tracking error changes only slightly. This indicates that LOGO's main benefit lies in preserving the currently executable interaction intent and keeping it consistent with the more distant goal, rather than merely improving how the motion looks.

Simulation results: three tasks and disturbance tests

The paper evaluates carrying, pushing, and kicking in a MuJoCo environment on a Unitree G1. Table 1 of the paper reports the success rate without disturbance (SR), the average error from the box to the goal at task end (Eobj, metres), and the success rate after the box is disturbed (SR-O) and after the robot is disturbed (SR-R).

The carry-box demonstrations in the training data cover only flat-ground carrying, with no demonstrations of carrying from one height to another. At evaluation, the box centre's starting height and target height are each set randomly between 0.15–0.8 metres; I-BFM achieves a 94.3% success rate under this cross-height carrying setting, showing generalisation to height combinations outside the training demonstrations. Supplementary tests also observed some adaptability to boxes of different sizes; there is currently no success-rate statistic broken down by box size.

Each item is 3 groups × 100 episodes, and "±" in the table is the sample standard deviation between groups; Eobj is the three-dimensional distance from the box to the goal at termination, counted in the mean at a maximum of 1 metre. In Carry Box, SR-O applies a 200 N downward external force to the box for 0.2 seconds; in Push/Kick, SR-O applies a 200 N horizontal external force to the box for 0.2 seconds; SR-R applies a 2000 N downward external force to the robot's pelvis for 0.3 seconds in all cases.

Push/Kick uses a 90-second limit for all methods, but the success criteria and some initial conditions still differ: I-BFM requires the box to stay within 0.20 metres of the goal for 100 control steps, Push/Kick judges by horizontal distance; OmniContact Push judges by the three-dimensional distance at the endpoint, and LessMimic Push follows its own initialisation. Comparisons should therefore also account for these protocol differences.

Real-robot demonstrations: one policy executing several kinds of interaction

The research team also demonstrated carrying, pushing, and kicking on a Unitree G1. The paper states that the real robot is located in a motion-capture space, and the system obtains the poses of the robot, box, and goal from external measurement; a box with 0.35-metre sides and 0.7 kilograms is used, and the policy runs at 50 Hz. The video shows continuous interaction between the robot and the box, as well as a retry after a failed carry.

Supplementary real-robot demonstrations also include a G1 waving and dancing, which can serve as a display of body motion diversity. These clips do not provide the control method or multi-episode test conditions.

What does this mean for research and engineering?

What is worth attention in I-BFM is that it puts "executing successfully" and "continuing after deviation" into the same control framework. The paper also shows that by updating the reward instruction, one policy can chain together multi-stage box manipulation. This gives multi-task robot systems a concrete direction: first learn reusable body–object interaction capability, then compose behaviours with task goals; whether a new task is feasible depends on the coverage of the learned behaviours and whether the reward design is good enough.

Problems to solve in the future

The next step is to take this kind of capability out of controlled experimental environments. First, onboard perception must replace the poses provided by motion capture, so that the states of the robot and the object can still be judged stably under occlusion and changing contact. Second, tasks must extend from carrying, pushing, and kicking to richer behaviours and object interactions such as sitting down, throwing a ball, and wiping a table, and generalisation across different objects and contact modes must be tested. Finally, the paper shows chaining actions by updating task instructions; higher-level planning still requires the robot to decide on its own which skills to use, in what order, based on goals and real-time feedback, and to adjust the plan after failure.