Replacing the plan or keeping it both fail
A humanoid pushing an unfamiliar cart cannot see its mass or how freely its rear casters swivel until it has already pushed. Those properties decide the shape of the maneuver: an omnidirectional cart can be driven sideways to its goal, while a constrained rear axle admits only a curved approach. The robot therefore has to plan, act, learn from the cart's response, and revise while the maneuver is still running.
The paper frames the difficulty as a tension. Regenerating the plan from every new estimate, on top of noisy pose estimates, disturbs the very trajectory the controller is following. Refusing to regenerate leaves the robot committed to a path its own measurements have ruled out. The question the authors ask is how much a new estimate should be allowed to change the plan.
The simulation results show both failure modes are real. Retaining the incorrect initial plan left pooled final position and yaw errors of 0.465 m and 0.277 rad, while repeatedly replacing it with a freshly generated plan (Fresh FM) produced even larger errors of 0.936 m and 1.149 rad. Neither extreme is a workable default.
Belief overlap sets how deep the revision goes
BOPI treats the plan as an encoding produced by a learned flow, a model that generates a timed cart trajectory when conditioned on the goal and on the current belief about the cart's physical properties. To revise, it runs that flow backward from the current plan under the stored belief, then forward again under the updated belief. The intermediate state is a trajectory encoding, not a physical pose the robot visits, and no fresh noise is added.
The depth of that backward run is not a tuned constant. It is set by the Bhattacharyya coefficient between the old and new belief distributions, which lies between zero and one. If the beliefs nearly agree, the reversal is shallow and the plan barely moves; as they diverge, more of the plan is rewritten. At full overlap both integrations have zero length and the plan is unchanged; at zero overlap the method performs a complete inversion and reconstruction.
The authors prove a bound on how far the encoded plan can move under exact integration: the change is capped by the smaller of a term proportional to the reversal distance and a term that grows with the difference between the two flow fields. Under the stated bounds, the change in the encoded plan approaches zero as belief overlap approaches one. This is the formal version of the design rule, not a measured result.
What the estimator knows after four seconds
The belief is produced by gated recurrent unit ensembles that predict rear-caster stiffness and cart mass from interaction history, in standardized log coordinates. Caster stiffness matters because small values let the rear wheels steer freely while large values approximate a locked rear axle. The estimator averages property estimates across models and represents each property as a Gaussian, which is what makes the analytic overlap between old and new beliefs computable.
After four seconds of interaction, the estimator distinguished free from constrained rear wheels with 91.5% to 98.4% accuracy and estimated cart mass with approximately 1 kg mean absolute error on held-out trajectories. These are estimator-level numbers on held-out trajectories, not end-to-end task success, and they set the quality of the belief that BOPI later acts on.
Beliefs change the shape and timing of paths
The planner is conditioned on the belief mean, so different beliefs produce different timed trajectories. Nearly free rear casters allow direct lateral motion and near-in-place rotation, whereas a constrained rear axle produces curved lateral paths and wider turning maneuvers. The plans also adapt their timing: at a cart mass of 18.3 kg, increasing rear-caster stiffness extended the lateral plan's duration from 5.11 to 9.42 s. Separately, raising the mass belief from 3.3 to 18.3 kg at a stiffness of 2000 increased lateral plan length by 8.5 cm.
These are properties of generated plans under specified belief settings, not measured robot outcomes. They matter because they show why a belief update is not a cosmetic correction: the same goal can require a different path geometry and a different schedule once the cart's properties are better known.
Smaller plan changes produce lower final error
The end-to-end simulation test used 18 goals with 32 seeds each, spanning one-meter translations, rotations of plus or minus 60 or 90 degrees ending at the starting position, combined translation and rotation, and one-meter translations with 45-degree turns under an incorrect mass belief. The first three categories incorrectly assumed freely steering rear wheels; the wrong-mass tasks used an omnidirectional cart with an assumed mass of 3.3 kg instead of 18.3 kg. Replanning ran at 5 Hz, and errors were computed from ground-truth poses against the fixed goal, averaged over each rollout's final two seconds and then across rollouts.
BOPI had the lowest error in every category, with pooled final position and yaw errors of 0.179 m and 0.152 rad. That is roughly 33% and 47% below fixed-depth replanning (RDM, at 0.267 m and 0.286 rad) and 39% and 56% below FlowEdit (0.295 m and 0.349 rad). The paper attributes this to BOPI making smaller changes between successive plans, producing consistent replanned trajectories as the robot progresses toward the goal.
A separate ablation isolates the inversion step. Holding the belief-overlap depth rule fixed but replacing flow inversion with forward noising, as in RDM, increased mean final position and yaw errors by 18.2% and 20.9% across the same 18 goals with 64 seeds. So the gain comes from both parts of the design: adapting the revision depth, and reversing the old plan through the learned flow rather than perturbing it with fresh noise.
A toy problem shows the same pattern
Before the cart task, the authors test the replanning rule on a nonlinear toy problem where beliefs about damping guide control generation. An incorrect damping estimate makes the initial plan unreliable: retaining it succeeds in only 2.4% of trials, compared with 97.2% when planning with the true damping. Each method was evaluated on 256 trials per flow across four independently trained flows, with success requiring a final position within 0.12 of the target and a speed below 0.15.
In that setting BOPI improved success over fixed-depth RDM by 14.8 plus or minus 10.4 percentage points, while reducing cumulative plan changes by 68.2% and final speed error by 45.1%. The toy problem uses different success criteria and a different physical quantity from the cart benchmark, so its numbers should not be read as cart-pushing performance; its role is to test the repair mechanism under controlled conditions.
Tracking the cart rather than a fixed offset
The whole-body controller tracks timed cart waypoints while preserving grasp and balance, and it is trained with reinforcement learning in simulation. Across masses and wheel constraints it followed cart waypoints with a mean RMS position error of 6.53 cm over 8 trials. Because the target is the object's path rather than a fixed body-to-cart offset, the robot can change its position relative to the cart during turning while still tracking the cart's waypoints.
This matters for reading the planning results: the replanner is judged on where the cart ends up, and the controller is what turns revised waypoints into joint targets. A rollout illustrated in the paper reached the goal with mean position and yaw errors of 0.036 m and 0.007 rad, which is a single example rather than an aggregate.
Where the evidence stops
The hardware demonstration is qualitative in the reported evidence: on a real non-holonomic cart, BOPI performs translation and turning with lower mean final position and tracking errors than the compared baselines, but the packet gives no hardware trial counts or error values. The quantitative comparison lives in simulation, in the mjlab G1-Dex3 environment with randomized posture and handle height between 0.78 and 1.00 m.
The scope is also planar cart pushing with wheeled objects. The authors name extension to three-dimensional goals and non-wheeled objects as a next step, so nothing here establishes how the overlap-based revision rule behaves outside that setting. The estimator's accuracy figures are likewise tied to four seconds of interaction and to the properties it was trained to infer.
The priority is proportional revision
The paper's distinctive claim is not that replanning helps or that estimation helps, but that the size of a revision should be governed by the size of the belief change. The comparison that carries this is the ablation: with the same overlap-based depth rule, reversing the old plan through the flow beat adding fresh noise by 18.2% in position error and 20.9% in yaw error, and with a fixed depth, BOPI beat RDM by about 33% and 47% on pooled final error.
For a reader deciding what to take from this, the practical split is between situations where successive beliefs largely agree and situations where they diverge. In the first, BOPI's rule leaves the plan nearly intact, which is exactly where aggressive replanning would have disturbed a maneuver that was already working. In the second, it allows a deep rewrite, which is where retaining the plan would have kept the robot on a path its measurements had ruled out. The rule needs no per-situation tuning, and its effect on the encoded plan vanishes as overlap approaches one.
Sources used in this answer
Belief-Overlap Partial Inversion: Object-Centric Flow Replanning for Humanoid Loco-Manipulation
BOPI scales mid-maneuver plan revision to the overlap between successive beliefs about cart mass and caster stiffness, lowering pooled final position and yaw error by roughly 33% and 47% versus fixed-depth replanning across 18 simulated goals.