Filter-Bank Observations for Low-Rate Limb Attitude Controllers

Learned controllers for acrobatic robots usually run slowly, at roughly 50 Hz. Robots that reorient in mid-air by swinging their legs, or that balance on a beam by counter-swinging their arms, are in this group. Their motion, however, carries information at much higher frequencies: limb whip, gearbox compliance, and structural ringing that the limbs themselves excite. Each control step, the policy sees one snapshot of the robot's state, and that snapshot is effectively a sample of a fast signal taken far below its Nyquist rate (twice its highest frequency). The fast content therefore does not disappear. It folds into the slow band and looks like genuine body motion. This folding is called aliasing. A simulator that reads out the exact state once per step and a real inertial sensor with its own internal filtering fold the fast content differently. Our proposal is that this mismatch in the sampling operator is an unmodeled cause of the hardware oscillations reported for leg-based free-fall reorientation. We propose to fix it with a result from sub-Nyquist sampling theory, rather than by raising the control rate.

The result comes from the theory of parallel filter-bank sampling. A signal passed through $M$ different pre-filters, each sampled at a low rate $f_s$, retains the information of a band roughly $M$ times wider than $f_s/2$. When bands are noisy, the information-optimal pre-filters suppress low signal-to-noise frequencies before sampling, rather than letting them fold onto clean ones (Chen et al.). We would give the low-rate policy a small bank of such filtered channels at every step instead of a single snapshot. Orientation, though, is not a vector space, so linear filtering of Euler angles or quaternion components is ill-defined near gimbal lock (the pitch-±90° configuration where two rotation axes line up) and across quaternion sign flips. We therefore build the bank on tangent-space rotation increments, which stay singularity-free.

Core Claim

For a fixed policy rate and a fixed number of observation channels per step, a pre-sampling filter bank has three properties. First, its filters are matched to the separation between rigid-body motion and structural vibration in the robot's spectrum. Second, it is applied to body-frame tangent increments of orientation. Third, it is used identically in simulation and on the target sensor model. We claim this bank will (a) reduce attitude-tracking error and high-frequency limb oscillation under sim-to-real transfer compared with snapshot observations and naive high-rate history stacking, and (b) remove a predicted non-monotonic dependence of control quality on policy rate.

Claim (b) is the sharpest test. Chen et al. proved that single-branch sub-Nyquist sampling without alias suppression has capacity that does not rise monotonically with sampling rate, because some rates land noisy bands on top of high-SNR ones. Multi-branch alias-suppressing banks remove these "wiggles." If the sampling-operator view is correct, the same wiggles should appear when policy performance is plotted against control rate for snapshot observations on a flexible robot, and should flatten under the filter bank. Neither the robotics literature nor the sampling literature has measured this.

Positioning Against Prior Work

The sampling side supplies the method. Chen et al. and Shannon Meets Nyquist model aliasing as a multi-input channel in which frequencies that fold together act like competing transmitters. They prove that water-filling pre-filters, which keep the best-SNR frequency from each folding set and null the rest, reach the capacity of the unsampled channel with $M$ branches at the Landau rate (the total width of the occupied bands). Kipnis and Chen et al. show that the capacity-optimal pre-filter is also the one that minimizes mean-squared reconstruction error for Gaussian sources. That equivalence lets us design filters from an estimation objective. Channel Capacity Under Sub-Nyquist Nonuniform Sampling shows that nonuniform or multi-rate sampling adds nothing over uniform sampling with SNR-optimal band selection. That result constrains the design: we do not pursue coprime multi-rate schemes (Huang et al.) and use uniform sampling at the policy rate throughout. All of this work assumes stationary Gaussian channels with known statistics and no feedback loop. We move the theory into a closed loop whose own actions shape the noise spectrum.

The robotics side motivates the problem. El-Agroudi et al. trained a PPO policy (proximal policy optimization, a standard on-policy reinforcement learning algorithm) in Isaac Sim. The policy uses five-bar-linkage legs as reaction masses to track 45° attitude steps during a 0.7 s free fall, and the authors attribute the oscillations seen on hardware to simplified actuator models and latency. We propose that part of that gap is the observation-sampling operator, and we design a test that separates it from actuator error. Sleiman et al. found an emergent tail effect: a quadruped's arm cancels base roll during trotting, with a 66.8% reduction in roll error, under 70 Hz model predictive control and 400 Hz whole-body control. Bai et al. used arm counter-swing and angular-momentum regularization to reach 90% success on a 25 cm beam with a Unitree H1-2. Both assume clean state estimates and never examine what their sensing folds into the control band.

The geometry side constrains the design. Schuck et al. feed reinforcement learning policies the logarithmic map of orientation into the tangent space and compose actions on the rotation group. This avoids gimbal lock and the quaternion double cover, where $q$ and $-q$ name the same rotation. Hussein et al. integrate attitude dynamics directly on the rotation group through the exponential map, so no drift off the manifold occurs. Neither addresses filtering an orientation signal over time. That is the operation we need, and it is where representation choice starts to matter.

Research Program

The signal being filtered is defined per control period $T$. Let $R(t)$ be the base orientation and $\omega(t)$ the gyroscope's body angular velocity. Within period $k$ we form the tangent increment

$$\xi_k(t) = \mathrm{Log}\big(R(kT)^{\top} R(t)\big) \in \mathbb{R}^3, \qquad t \in [kT - LT,\, kT],$$

where $L$ is a short look-back measured in control periods and Log maps a rotation to its axis-times-angle vector. $\xi_k$ is a smooth vector-valued signal as long as the body rotates by less than 180° over the window. At the roughly 110°/s rates reported by El-Agroudi et al. and a 20 ms period, the body turns about 2° per step, far from that boundary. Each observation channel is

$$y_{m}[k] = (h_m * s)(kT), \qquad m = 1, \dots, M,$$

where $s$ stacks $\xi_k$, $\omega$, and joint velocities, and $h_m$ are pre-filters applied to the sensor's internal high-rate stream before decimation to the policy rate. "Pre-sampling" here means before the policy's sampling. This lets the bank be implemented digitally after an oversampling sensor, which is how modern inertial sensors already work.

The filters are designed in three steps. First, we estimate the power spectra of the rigid-body component and the structural and actuator-vibration component. We do this by running the reference policy in a simulator with flexible links and compliant gear models, and by recording pole-rig data where hardware is available. Second, for each set of frequencies that fold onto the same baseband frequency at rate $1/T$, we apply the water-filling rule of Chen et al.. The $M$ branches are assigned the $M$ highest-SNR members of each set, where "signal" means rigid-body motion relevant to attitude, and all remaining members are nulled. Third, the policy is trained with identical filters in simulation, and the sensor model reproduces the target hardware's internal sampling chain.

We will build the flexible simulation testbed, the bank-design procedure, and policies for free-fall reorientation modeled on the Eurepus task. The target set extends beyond 45° steps to large pitch reorientations that pass through ±90°, which is exactly where Euler-angle filtering breaks. As a secondary test of generality, we add arm-as-tail roll stabilization in a trotting quadruped modeled on the Sleiman et al. setting.

Baselines

The naive reference is the observation design used by El-Agroudi et al.: one state snapshot per policy step. We retrain it on our testbed with their reported setup of PPO in Isaac Sim with domain randomization. That paper reports tracking of 45° steps in all axes and an average base angular velocity of 110°/s. We found no reported numerical attitude-tracking error on Lacuna for this task, so our reproduced snapshot policy fixes the number to beat. Beating it means lower geodesic attitude error at the end of the fall and lower joint-torque power above the policy Nyquist frequency when transferred to the flexible, realistic-sensor simulator.

The strongest existing practical competitor is oversample-and-decimate. The sensor stream passes through one good anti-aliasing low-pass filter and is decimated to the policy rate, which is standard sensor practice. It suppresses folding but discards everything above $f_s/2$, so it isolates the multi-branch half of our claim. It receives the same simulator, training budget, and hyperparameter search as our method.

The second strong competitor is high-rate history stacking: the last $M$ raw high-rate samples, which amount to a delay-line filter bank with no alias suppression. A related variant passes a longer raw history through a learned one-dimensional convolutional encoder, which is a filter bank learned end-to-end. Both are matched to our method in channels per step (for the raw stack) or in parameters and training steps (for the learned encoder).

A third reference raises the policy rate until snapshot observations stop aliasing. This is the upper bound a skeptic would cite, and it is charged its real cost in added latency and compute per step.

Two ablations isolate the key ingredients. The first keeps the same filter bank but applies it to Euler angles or to quaternion components, which isolates the tangent-increment construction. The second applies a designed bank on hardware and in the test simulator but uses snapshots during training, which isolates sim-to-real matching of the sampling operator from the bank itself.

Why not just…?

Why not just run the policy faster? If a 500 Hz policy with snapshots matches the bank at 50 Hz, the practical case weakens. The rate sweep in our experiments measures this directly. Our expectation rests on two points. Faster policies are harder to train and to deploy on shared buses. More importantly, the snapshot curve should be non-monotonic in rate, by analogy with the capacity wiggles in Chen et al., so "faster" is not reliably "better." If the sweep turns out monotone and a moderate rate suffices, we will report that as a negative result for the aliasing hypothesis.

Why not just low-pass filter and decimate, as every sensor already does? This is a real threat to the alias-suppression half of the idea. It already prevents folding, and the file's claim that digital filtering cannot help applies only after sub-Nyquist sampling, not to an oversampling sensor. Our method differs only in what it keeps. A single low-pass branch throws away limb and structural dynamics that may carry attitude-relevant information, such as a whipping leg's contribution to base momentum. The comparison against the $M=1$ low-pass baseline settles whether the extra branches earn their place. If they do not, the contribution reduces to the sampling-operator-matching result.

Why not just stack raw high-rate history, or learn the filters with a convolutional encoder? A learned encoder can in principle discover the water-filling bank, so this is the most serious threat. Our bet is that the designed bank helps most in low data and under randomized resonance frequencies, where the learned encoder can overfit to one folding pattern. We test designed filters, learned filters, and designed filters used as initialization for learning. If the learned encoder matches in sample efficiency and robustness, we say so. The remaining contribution is then the diagnosis, a principled channel budget $M$ derived from the spectrum, and the tangent-space construction that any learned bank also needs.

Why not just use quaternions or a full rotation matrix, since gimbal lock is solved? As single-step inputs, these are fine, and Schuck et al. show Lie-algebra inputs already train faster than Euler angles, converging in about 18.8 s versus about 45.1 s. The problem appears when one filters across time. Averaging quaternion components mixes $q$ and $-q$ unless signs are tracked, and averaging rotation matrices leaves the rotation group. Tangent increments relative to the current sample make the linear theory valid by construction. The representation ablation on trajectories through ±90° pitch shows whether this matters in practice or only on paper.

Proposed Experiments

The primary testbed is a Eurepus-like free-fall reorientation task in Isaac Sim. The robot has four five-bar legs and a 0.7 s fall, with targets extending from the reported 45° steps to 90° and 150° reorientations. Training uses a rigid-body simulator. Evaluation uses a second simulator that adds link flexibility, gearbox compliance, and a sensor chain that samples internally at kHz rates with configurable internal filtering. Domain randomization covers resonance frequencies and damping.

The metrics are four. The first is geodesic attitude error at touchdown and averaged over the last 0.2 s of the fall. The second is success within a 10° tolerance. The third is joint-torque spectral power above the policy's Nyquist frequency, as an oscillation index. The fourth is training samples needed to reach a fixed success rate.

The first decisive check is the rate sweep. Snapshot and filter-bank policies are trained at policy rates from 25 to 200 Hz, and success is plotted against rate. The sampling-operator hypothesis predicts dips in the snapshot curve at rates where a dominant structural mode folds onto the rigid-body band, and a flat curve for the bank. The absence of such dips in the snapshot curve would falsify the central mechanism.

The second decisive check is the matching ablation. Holding actuator models fixed and varying only the observation sampling operator between train and test separates this effect from the actuator-model gap that El-Agroudi et al. blame for their oscillations.

The third check is the channel-budget curve, which varies $M$ from 1 to 8 at fixed rate. It should saturate near the number of attitude-relevant bands in the measured spectrum, as the Landau-rate result predicts, and not keep improving the way raw history stacking might.

The secondary test repeats the transfer comparison on quadruped arm-as-tail roll stabilization, modeled on Sleiman et al., where their reported 66.8% roll-error reduction relative to a rigid arm is the reference effect size. A pole-rig hardware check, following the rotating-pole protocol of El-Agroudi et al., is included if a comparable platform is available. It is not a precondition for the simulation results.