Unknown mass turns a nominal model into a wrong model
A humanoid lifting a box it has not weighed does not merely face an extra downward force. The paper notes that an unknown payload simultaneously perturbs the robot's mass, center-of-mass location, and inertia tensor, so the mismatch between the controller's model and the real system is a coupled, state-dependent nonlinearity rather than a simple bounded disturbance. That distinction matters because it explains why the authors did not simply add a disturbance observer or a robust control law: those routes would need an explicit disturbance model, a reliable uncertainty bound, or online identification of the payload, quantities the paper describes as noisy on a full-scale humanoid.
The architecture therefore splits the problem. A centroidal model predictive controller (MPC) plans center-of-mass motion and contact forces that are dynamically feasible, and a whole-body control layer tracks those references in joint space under kinematic and dynamic constraints. Neither layer, on its own, knows how heavy the object is. The adaptive fuzzy balance controller is the piece that reacts to the resulting tracking error and produces compensation torques, which are then distributed across the supporting-leg joints through a torque-mapping step. The paper's framing is that this preserves the constraint-aware, optimal structure of the MPC while adding a reactive correction layer.
Compensation is mapped to the stance leg, not the arms
The adaptive fuzzy controller estimates a torque compensation from tracking errors and augments the nominal MPC output. The paper describes the distribution step as an optimization: the leg joint torques are chosen to minimize a quadratic cost subject to the requirement that they reproduce the desired task-space compensation torque at the pelvis, treated as the end effector during the stance phase. In plain terms, the controller decides how much correction is needed at the pelvis, then asks which combination of supporting-leg joint torques delivers it most cheaply.
This is where the design choice becomes concrete. The correction is not applied as an arbitrary whole-body push; it is routed through the stance leg, the part of the robot already in contact with the ground and able to convert joint torque into a balance-relevant force at the pelvis. The paper's stated motivation for fuzzy logic over model-based alternatives is that the payload-induced mismatch is nonlinear and state-dependent, while conventional fuzzy controllers are described as relying on heuristic rule design and lacking predictive optimization and constraint handling. The proposed combination is meant to keep MPC's constraint handling and add fuzzy logic's tolerance for unmodeled variation.
A triple-integrator planner makes posture shifts smooth and feasible
Lowering the pelvis to grasp an object and standing back up are not small perturbations. The paper models the standing robot as a three-link planar chain with hip, knee, and ankle angles and generates trajectories with a discrete-time triple-integrator model, where the input is jerk. The optimization minimizes weighted tracking error to a reference joint configuration plus the size of the jerk input, subject to the integrator dynamics, joint position, velocity, acceleration and jerk bounds, and a constraint that the center-of-mass projection stays inside the support polygon.
A separate mechanism chooses where the pelvis sits horizontally. Given a desired pelvis height and pitch, inverse kinematics can have infinitely many horizontal solutions that keep the CoM projection inside the support polygon, so the algorithm steps the horizontal position and returns the first configuration whose joint angles fall within limits. The paper reports posture-transition examples in which the CoM projection stays inside the support polygon throughout and the pelvis x-position shifts between initial and final poses because of this selection step. This is the proactive half of the strategy: rather than waiting for a balance error, the planner chooses a posture from which the load is easier to hold.
The clearest comparison is a collapse versus a maintained margin
The most direct evidence for combining the modules comes from MuJoCo simulation. With a standard centroidal MPC and no adaptive balancing control, the paper reports that uncompensated CoM shifts, compounded by non-optimized kinematic trajectories, pushed the robot past its stability margin and caused it to tip forward and collapse in stages 1 to 3 while holding an unknown heavy payload. With the proposed framework, the robot computed compensation torques and synthesized feasible waypoints that proactively lowered its center of mass, keeping the CoM within the stability margin through the manipulation cycle.
This is a qualitative comparison rather than a reported numeric margin, and it should be read that way: the simulation isolates the contribution of adaptation plus posture planning against a non-adaptive baseline, but the packet does not give a quantified stability-margin value for either condition. What it does establish is the failure mode the design targets, and that the two components were used together in the successful case.
Heavier loads raise control effort and produce small oscillations
On the 28-degree-of-freedom humanoid, the control stack runs at three rates: the centroidal MPC at 50 Hz, the adaptive framework at 500 Hz, and the low-level torque controller above 1 kHz. In the 7 kg lifting trial, the robot began lowering its pelvis and leaning forward at about 30 s, completed the sitting motion within 3 s, grasped the object, stood up between 55 s and 80 s, and then placed the object back down, with smooth posture and joint-state transitions reported across pelvis height, pitch, and hip, knee and ankle angles and torques.
The 10 kg trial is the more informative one for payload uncertainty. The robot moved into a bowing posture at about 20 s and grasped the object around 30 s, then lifted it while gradually recovering torso pitch. Because the payload mass changed substantially and torso dynamics shifted with it, the adaptive controller increased its control effort, which the paper reports as small oscillations in pelvis height and pitch during the initial lifting phase; once pitch stabilized, the robot recovered height and returned to the nominal upright position. The oscillations are a visible cost of the reactive correction, not a failure, and they appear specifically in the heavier-load condition.
The same fuzzy tuning carried over to walking
The second hardware scenario chained the full task: bow and grasp at 83 s, lift and return to the home posture at 110 s, walk while carrying the 7 kg load for 15 s from 140 s to 155 s, then place the object and recover. The paper reports that the robot walked normally and maintained stable locomotion under the uncertain payload.
One detail deserves attention because it speaks to how general the compensation is. The fuzzy parameters were not re-tuned for walking, which the authors present as evidence that the tuning is not lifting-specific. That is a claim about parameter transfer within this robot and these trials, not a demonstration that the controller transfers to other robots, other payload shapes, or faster locomotion; the packet provides no such test.
Reactive fuzzy correction remains the framework's weak point
The paper's own discussion identifies the fuzzy logic as heuristic and reactive, and states that future work will pursue learning-based methods to optimize compensation strategies and proactive disturbance observers to improve agility in highly dynamic environments. The limitations section of the extracted text is not readable in the packet, so the specific enumerated limitations cannot be reported here. What can be said is that the quantitative record is thin in one respect: the hardware results are presented as time histories of pelvis height, pitch, joint angles and torques for the 7 kg, 10 kg, and lifting-walking cases, and the abstract's claim of improved balance robustness and reduced tracking degradation relative to non-adaptive baselines is supported in the packet mainly by the simulation collapse comparison rather than by a reported error metric.
The practical reading is therefore a priority ordering rather than a blanket endorsement. The strongest evidence is for the combination of proactive posture planning and reactive leg-torque compensation under a payload the controller has not identified, at the tested masses of 7 kg and 10 kg and in the tested sit-to-stand, lift, carry and place sequence. The weakest link the authors themselves name is that the corrective layer reacts after error appears; if a task demands faster disturbance rejection, the paper points toward learned compensation and proactive observers rather than more tuning of the existing fuzzy rules.
Sources used in this answer
Adaptive Whole-Body MPC with Optimal Posture Planning for Humanoid Robots under Unknown Payloads
A centroidal MPC, an adaptive fuzzy balance controller distributing corrective torques to supporting-leg joints, and a triple-integrator waypoint posture planner together kept a humanoid balanced while lifting 7 kg and 10 kg payloads and walking with 7 kg, where a non-adaptive centroidal MPC collapsed in simulation under an unknown heavy payload.