When a humanoid walks a narrow beam or a legged robot is in mid-air, it can barely push on the ground to stay upright. Its main remaining tool is to swing its own limbs: moving an arm one way rotates the body the other way, because total spin is conserved. An arm has limited travel, though. Once it reaches a joint limit it can no longer correct anything in that direction, and every bit of arm motion spent on a task such as holding a box is travel that cannot be used for balance. Current controllers settle this trade-off with fixed rules. They either freeze the arms, permanently dedicate them to balance, or use hand-tuned cost weights that never change during a run. Cloud data centers made the same mistake with fixed utilization thresholds. Admission rules that bound risk using the mean and variance of demand improved on those thresholds.
This note proposes treating each limb's remaining capacity to correct attitude, its "momentum reserve," as a resource that manipulation requests must be admitted against. A manipulation command gets the arm only if a one-sided Chebyshev (Cantelli) bound, computed from the recent mean and variance of the robot's balance demand, says the leftover reserve covers disturbances with probability at least $1-\varepsilon$. The claim is that this variance-aware rule beats every fixed allocation on the trade-off between task success and falls. It should also make the realized fall rate roughly controllable through the single parameter $\varepsilon$, which no weight-tuning scheme offers.
Core Claim
The proposal combines one idea from each field. From robotics comes the fact that limbs act as internal reaction masses. Four results establish this. Sleiman et al. find that lowering the arm's motion penalty in a whole-body model predictive controller turns a quadruped's manipulator into a balancing "tail" and cuts base roll error by 66.8%. Bai et al. show that freezing the upper body sharply degrades humanoid crossing of a 25 cm path. Ficht et al. use the torso itself as a flywheel. El-Agroudi et al. reorient a falling quadruped using only its legs. All four motivate the proposal. In each of them, however, the split of limb capacity between balance and other duties is fixed before the run starts. The file also names the open problem: limbs saturate at joint stops and then need uncoordinated "momentum dumping."
From systems comes the admission rule. Dierks, Kash, and Seuken show that raw utilization is a poor proxy for risk. A quiet cluster at 60% is safer than a bursty one at 40%. Replacing a fixed threshold with Markov- and Cantelli-bound admission raised utilization from about 50% to 66–67% at controlled violation risk. We reuse their method and port its central lesson: "how close is the arm to its joint limit" plays the role of raw utilization, and is just as poor a proxy for fall risk as utilization was for SLA risk.
Concretely, let $R(q)$ be the reserve: the largest base-attitude correction, about the critical axis (roll on a beam), that the limbs can still produce from configuration $q$ within a short horizon $T$, given joint limits and torque limits. It is computed from the model's centroidal inertia, using the momentum-conservation relation in the file. Let $D$ be the corrective demand that disturbances will place on the robot over the same horizon. Let $\mu_D$ and $\sigma_D$ be its mean and standard deviation, estimated online from a sliding window of proprioceptive history (IMU readings and the estimated centroidal angular momentum). A manipulation request that would consume reserve $d$ is admitted when
$$\mu_D + \sqrt{\tfrac{1-\varepsilon}{\varepsilon}}\,\sigma_D \;\le\; R(q) - d .$$
The square-root factor comes directly from Cantelli's inequality, $P(D-\mu_D \ge k\sigma_D)\le 1/(1+k^2)$, solved for $k$. Dropping the variance term gives the first-moment (Markov-style) variant. If the inequality fails, the request is scaled down to the largest admissible $d$ rather than rejected outright, so the arm's "role" shifts continuously between task and balance. This is the physical analogue of the GPU role switching in EPD disaggregation.
What would be new is the conserved, saturating quantity being budgeted (internal momentum capacity rather than torque or compute), the task-level admission framing, and the calibration claim tied to $\varepsilon$. This is narrower than the file's broad "dynamic reallocation" gap, and we think it is more defensible.
Research Program
We first build a low-level controller conditioned on an arm-allocation input $w\in[0,1]$. Here $w=0$ means the arms serve only balance, and $w=1$ means they track the manipulation reference as closely as possible. The controller is a proprioceptive PPO policy (proximal policy optimization, a standard RL algorithm) trained across randomly sampled $w$. It uses the reward structure of DBHL from Bai et al., namely the Zero Moment Line reward, angular momentum regularization, and multiplicative action noise. It is trained in Isaac Gym on that paper's 20-level curriculum, which narrows the path from 1.0 m to 0.2 m, with an added task: holding a carried object at a target pose with both hands. Conditioning on $w$ is what makes the comparison clean. Every allocation method, whether static, heuristic, or admission-based, drives the same trained policy and differs only in how it sets $w$ over time.
Second, we compute $R(q)$ in closed form from the simulator's mass matrix and joint limits. We validate it in the one regime where momentum conservation is exact: free fall of a quadruped, as in El-Agroudi et al.. There we check that the predicted reserve matches the largest reorientation actually achieved before joint limits bind, including the case where one leg is held back for touchdown.
Third, we build the demand estimator and the admission layer. We study how window length trades off responsiveness against the non-stationarity weakness that Dierks et al. report for moment-based rules.
Baselines
The naive reference point is the frozen-arm configuration, $w$ held at the task value with no balance assistance. This reproduces the upper-body-locked ablation of Bai et al., which they report degrades balance severely. It anchors the floor.
The strongest existing method is the static allocation of Sleiman et al.: one fixed trade-off weight, given the same tuning budget as our method. We sweep $w$ over a dense grid on the same terrain and disturbance seeds, so that the static baseline is represented by its entire Pareto front, not one hand-picked point. Their reported figure is a 66.8% reduction in base roll error relative to a rigid arm, on ANYmal C with a DynaArm. That robot is not in our benchmark suite, so beating them here means our admission curve lies strictly outside the best static front: higher task success at equal fall rate on held-out disturbance schedules.
The end-to-end competitor is a DBHL-style policy that chooses its own arm usage, trained with the same sample budget and reward. On hardware DBHL reports a 90% success rate on a 25 cm path, against 0% for a baseline RL policy Bai et al.. No reported result was found on Lacuna for any method on narrow-path crossing while holding an object, so all numbers for this combined task will be produced by us.
The ablations isolate the key ingredient. A heuristic switch drops $w$ when roll rate crosses a threshold. A distance-to-joint-limit rule uses raw "utilization" in place of $R(q)$. A mean-only admission rule removes $\sigma_D$. If the mean-only rule matches the Cantelli rule, the variance term is not doing the work. If the joint-limit rule matches it, the reserve model is not doing the work.
Why not just…?
Why not just tune a fixed arm weight well? This is the real threat. In stationary conditions a well-chosen constant may sit on or near the admission curve. Our sweep therefore gives the static baseline its full front. The proposal predicts an advantage only when disturbance variance changes within a run, for example when calm beam segments alternate with push-heavy ones. We will report stationary and non-stationary schedules separately. If the static front matches ours under non-stationary schedules too, the claim is falsified.
Why not just let an end-to-end RL policy learn the arbitration? Flat policies can produce whole-body recovery spontaneously, as Ferigo et al. show for push recovery over 23 joints. However, HumanoidBench finds flat agents struggle once manipulation is added to the action space. More importantly, a learned policy offers no handle like $\varepsilon$ for setting acceptable fall risk after training. The end-to-end baseline tests the first point directly. The calibration experiment tests the second.
Why not just use chance-constrained or robust model predictive control? Mathematically, the Cantelli rule is a distribution-free chance constraint, and we do not claim otherwise. The difference lies in what is constrained and where. We constrain a task-level budget over a reserve computed in closed form, sitting above any low-level controller, at negligible compute cost. A full chance-constrained MPC over whole-body dynamics would face the real-time burden that already holds Sleiman et al. to 70 Hz. If compute allows, a chance-constrained version that uses the same moment estimates would be a welcome extra comparison.
Proposed Experiments
All main experiments run in simulation on the H1 humanoid model, in Isaac Gym, with MuJoCo cross-checks following Bai et al.. The task is a beam-carry: crossing paths 0.2–0.5 m wide while holding an object, under randomized lateral pushes. The push schedules are drawn from stationary distributions and from regime-switching distributions whose variance changes partway along the beam. A second suite uses the HumanoidBench H1 balance and manipulation environments with the hands fixed, which avoids the dexterity-tax confound HumanoidBench. The free-fall quadruped test validates $R(q)$.
We measure fall rate, object pose-tracking error, task completion, and the number of joint-limit saturation events, all on held-out seeds.
Three checks decide the question. First, does the Cantelli admission curve dominate the best static Pareto front under regime-switching disturbances? Second, does realized fall probability track the target $\varepsilon$ across the range 0.01–0.2, with Cantelli remaining conservative where the mean-only rule is not? Third, mirroring Dierks et al., is distance to joint limits a weaker predictor of falls than the variance-adjusted reserve margin? We will measure this as area under the curve for predicting falls from each signal. A negative answer on the first or third check would mean the systems analogy adds nothing beyond better weight tuning, and we would report that outcome.
References
- A Unified MPC Framework for Whole-Body Dynamic Locomotion and Manipulation
- Humanoid Whole-Body Locomotion on Narrow Terrain via Dynamic Balance and Reinforcement Learning
- Direct Centroidal Control for Balanced Humanoid Locomotion
- In-Flight Attitude Control of a Quadruped using Deep Reinforcement Learning
- On the cluster admission problem for cloud computing
- Efficiently Serving Large Multimodal Models Using EPD Disaggregation
- On the Emergence of Whole-Body Strategies From Humanoid Robot Push-Recovery Learning
- HumanoidBench: Simulated Humanoid Benchmark for Whole-Body Locomotion and Manipulation