Proprioceptive Scenario Gating for Legged Robots

• Proprioceptive scenario gating is a probabilistic estimation strategy that fuses multiple contact hypotheses to improve the accuracy of legged-robot odometry.
• It employs an interacting multiple model (IMM) framework to dynamically adjust confidence between rolling-contact and slip modes by inflating process noise.
• The method achieves significant drift reduction—up to a 70% improvement in state estimation—while maintaining real-time performance on standard computational platforms.

Searching arXiv for the specified paper and closely related context papers. arxiv_search(query="(Li et al., 31 Mar 2026) Interacting Multiple Model Proprioceptive Odometry for Legged Robots", max_results=5, sort_by="submittedDate") Proprioceptive scenario gating, in the form presented for legged-robot odometry, is an interacting multiple model (IMM)-based estimation strategy that represents contact uncertainty through multiple proprioceptive contact hypotheses and fuses them online within a unified probabilistic framework. In "Interacting Multiple Model Proprioceptive Odometry for Legged Robots" (Li et al., 31 Mar 2026), the framework is designed for conditions in which exteroceptive sensors are unreliable or degraded and state estimation must rely primarily on IMU, joint-encoder, and contact-related leg-kinematics information. The central premise is that most proprioceptive odometry methods assume idealized point contact, whereas real locomotion frequently violates that assumption; the proposed estimator therefore gates between rolling-contact and slip hypotheses rather than committing to a single contact model.

1. Problem setting and conceptual basis

State estimation for legged robots is described as challenging because legged odometry generally suffers from limited observability and depends critically on measurement constraints to suppress drift. When exteroceptive sensing is unavailable or degraded, these constraints are derived mainly from proprioceptive measurements, particularly contact-related leg kinematics information. The specific difficulty addressed is that existing proprioceptive odometry methods often rely on an idealized point-contact assumption, which may be violated during locomotion; when that occurs, the effectiveness of proprioceptive constraints can be significantly reduced, degrading estimation accuracy (Li et al., 31 Mar 2026).

Within this formulation, proprioceptive scenario gating is realized by maintaining multiple contact-mode hypotheses for each stance foot and assigning them time-varying probabilities. The method does not replace proprioceptive odometry with exteroceptive correction; instead, it adapts the internal trust placed in proprioceptive constraints under different contact conditions. A plausible implication is that "scenario gating" here is best understood as mode-conditioned probabilistic filtering over contact regimes rather than discrete logic-based mode switching.

2. Contact hypotheses and mode-dependent uncertainty

The framework considers two contact-mode hypotheses for each stance foot:

• Rolling-Contact Mode $(i = 1)$: the foot rolls on the ground without significant slip.
• Slip Mode $(i = 2)$: foot-ground contact is unstructured and partially slipping.

Under both modes, the nominal continuous-time dynamics are the same: an IMU-driven body with foot-velocity augmentation. The difference between modes appears only in the process noise associated with the foot-velocity states. Specifically,

$Q_{v_f}^{(1)} = Q_{v_f}$

for the rolling-contact mode, corresponding to nominal, tight foot-velocity confidence, whereas

$$Q_{v_f}^{(2)} = \alpha \cdot Q_{v_f}, \quad \alpha > 1$$

for the slip mode, corresponding to inflated covariance that represents slip-induced uncertainty (Li et al., 31 Mar 2026).

This construction is technically significant because the framework does not define separate nominal rigid-body dynamics for rolling and slipping. Instead, it retains a common dynamical model and expresses contact uncertainty through mode-dependent stochasticity in the foot-velocity block. This suggests that the method treats slippage primarily as uncertainty inflation in contact-consistent foot motion, rather than as a fully distinct kinematic regime with a separate deterministic evolution law.

3. State-space formulation and rolling-aware proprioceptive constraints

The full error-state Kalman filter (ESKF) state is defined as

$$x = [p, v, G, f_1,\ldots,f_4, v_{f1},\ldots,v_{f4}, b_a, b_\omega]^T,$$

where $p \in \mathbb{R}^3$ and $v \in \mathbb{R}^3$ are body position and velocity, $G \in SO(3)$ is body orientation, $f_l \in \mathbb{R}^3$ and $v_{f,l} \in \mathbb{R}^3$ are contact-foot position and velocity for leg $(i = 2)$0, and $(i = 2)$1 are accelerometer and gyroscope biases.

For each mode $(i = 2)$2, the discrete-time state-transition model is

$(i = 2)$3

with IMU input

$(i = 2)$4

Using fourth-order Runge-Kutta on the continuous dynamics, the linearized ESKF form is

$(i = 2)$5

with

$(i = 2)$6

where $(i = 2)$7 and $(i = 2)$8.

The measurement model is explicitly described as a rolling-aware kinematic update:

$(i = 2)$9

Crucially, both modes use the same rolling-aware measurement function; there is no separate slip measurement map. For each stance foot $Q_{v_f}^{(1)} = Q_{v_f}$0,

$Q_{v_f}^{(1)} = Q_{v_f}$1

where $Q_{v_f}^{(1)} = Q_{v_f}$2 is computed from the leg-kinematic Jacobian and $Q_{v_f}^{(1)} = Q_{v_f}$3 is the known foot radius vector. The residuals enforce body-foot relative pose consistency, body velocity relative to foot velocity, and foot-end velocity consistency with pure rolling (Li et al., 31 Mar 2026).

A common misunderstanding would be to interpret the slip mode as requiring a different observation function. The formulation explicitly states the opposite: the two modes share the same rolling-aware measurement function, and the distinction lies in the process noise assigned to foot-velocity states. This design couples contact-awareness to estimator confidence rather than to a mode-specific sensor model.

4. IMM machinery and probabilistic scenario gating

The gating mechanism is implemented through a two-mode Markov chain with transition matrix

$Q_{v_f}^{(1)} = Q_{v_f}$4

These values can be hand-tuned to reflect expected average slip duration or learned offline by counting mode-switch statistics in representative data.

Given mode posteriors $Q_{v_f}^{(1)} = Q_{v_f}$5 at time $Q_{v_f}^{(1)} = Q_{v_f}$6, the IMM recursion proceeds in five stages.

First, the filter mixes priors. The predicted mode weights are

$Q_{v_f}^{(1)} = Q_{v_f}$7

and the mixing probabilities are

$Q_{v_f}^{(1)} = Q_{v_f}$8

The mixed initial state for mode $Q_{v_f}^{(1)} = Q_{v_f}$9 is

$$Q_{v_f}^{(2)} = \alpha \cdot Q_{v_f}, \quad \alpha > 1$$0

with mixed covariance

$$Q_{v_f}^{(2)} = \alpha \cdot Q_{v_f}, \quad \alpha > 1$$1

Second, each mode performs a mode-conditioned prediction:

$$Q_{v_f}^{(2)} = \alpha \cdot Q_{v_f}, \quad \alpha > 1$$2

Third, each mode applies a mode-conditioned update using

$$Q_{v_f}^{(2)} = \alpha \cdot Q_{v_f}, \quad \alpha > 1$$3

followed by

$$Q_{v_f}^{(2)} = \alpha \cdot Q_{v_f}, \quad \alpha > 1$$4

$$Q_{v_f}^{(2)} = \alpha \cdot Q_{v_f}, \quad \alpha > 1$$5

Fourth, the mode likelihood is computed as

$$Q_{v_f}^{(2)} = \alpha \cdot Q_{v_f}, \quad \alpha > 1$$6

where $$Q_{v_f}^{(2)} = \alpha \cdot Q_{v_f}, \quad \alpha > 1$$7 is the dimension of $$Q_{v_f}^{(2)} = \alpha \cdot Q_{v_f}, \quad \alpha > 1$$8.

Fifth, the mode posterior is updated:

$$Q_{v_f}^{(2)} = \alpha \cdot Q_{v_f}, \quad \alpha > 1$$9

The final estimate is then fused across modes:

$$x = [p, v, G, f_1,\ldots,f_4, v_{f1},\ldots,v_{f4}, b_a, b_\omega]^T,$$0

$$x = [p, v, G, f_1,\ldots,f_4, v_{f1},\ldots,v_{f4}, b_a, b_\omega]^T,$$1

This recursion is the operational core of scenario gating. It performs neither hard contact labeling nor post hoc trajectory selection; instead, it executes online probabilistic fusion whose weights are driven by residual consistency under each contact hypothesis.

5. Empirical performance and robustness characteristics

The reported simulations and real-world experiments indicate that explicitly modeling both rolling and slip improves pose estimation accuracy over single-model filters while maintaining comparable computational efficiency (Li et al., 31 Mar 2026). The numerical results are organized below.

Setting	Reported result
Simulation, straight-line, $$x = [p, v, G, f_1,\ldots,f_4, v_{f1},\ldots,v_{f4}, b_a, b_\omega]^T,$$2	IEKF ATE_pos $$x = [p, v, G, f_1,\ldots,f_4, v_{f1},\ldots,v_{f4}, b_a, b_\omega]^T,$$3; ESKF-R $$x = [p, v, G, f_1,\ldots,f_4, v_{f1},\ldots,v_{f4}, b_a, b_\omega]^T,$$4; IMM-PO $$x = [p, v, G, f_1,\ldots,f_4, v_{f1},\ldots,v_{f4}, b_a, b_\omega]^T,$$5
Slippery patch, $$x = [p, v, G, f_1,\ldots,f_4, v_{f1},\ldots,v_{f4}, b_a, b_\omega]^T,$$6	IEKF ATE_pos $$x = [p, v, G, f_1,\ldots,f_4, v_{f1},\ldots,v_{f4}, b_a, b_\omega]^T,$$7; IMM-PO $$x = [p, v, G, f_1,\ldots,f_4, v_{f1},\ldots,v_{f4}, b_a, b_\omega]^T,$$8
Real complex terrain, $$x = [p, v, G, f_1,\ldots,f_4, v_{f1},\ldots,v_{f4}, b_a, b_\omega]^T,$$9	IEKF ATE_pos $p \in \mathbb{R}^3$0; IMM-PO $p \in \mathbb{R}^3$1

The straight-line simulation is summarized as IEKF ATE_pos $p \in \mathbb{R}^3$2 ESKF-R $p \in \mathbb{R}^3$3 IMM-PO $p \in \mathbb{R}^3$4, which is described as an approximately $p \in \mathbb{R}^3$5 further reduction beyond the rolling-aware baseline. In the slippery-patch experiment, IEKF ATE_pos $p \in \mathbb{R}^3$6 IMM-PO $p \in \mathbb{R}^3$7, corresponding to a $p \in \mathbb{R}^3$8 reduction. In real complex terrain over $p \in \mathbb{R}^3$9, IEKF ATE_pos $v \in \mathbb{R}^3$0 IMM-PO $v \in \mathbb{R}^3$1, corresponding to a $v \in \mathbb{R}^3$2 reduction.

The reported mode behavior is that mode probabilities correctly latch onto slip when foot residuals grow and revert to rolling otherwise, yielding both bias reduction through the rolling model and drift suppression through the slip model. This suggests that the estimator’s benefit is not only numerical averaging across models, but also targeted relaxation of overly strict proprioceptive constraints when residual structure becomes inconsistent with pure rolling.

6. Computational profile, tuning, and limitations

The implementation notes report runtime per step of approximately $v \in \mathbb{R}^3$3; for $v \in \mathbb{R}^3$4, the runtime is approximately $v \in \mathbb{R}^3$5 on a $v \in \mathbb{R}^3$6 CPU, and the system remains real-time (Li et al., 31 Mar 2026). This computational profile is consistent with the choice to use only two hypotheses and to keep their nominal dynamics and observation structure shared.

Three tuning parameters are highlighted. First, $v \in \mathbb{R}^3$7, the foot-velocity noise scale in slip mode, trades between over-confidence and under-confidence. Second, $v \in \mathbb{R}^3$8 should be set with high self-transition, for example $v \in \mathbb{R}^3$9-$G \in SO(3)$0, to avoid jitter. Third, IMU, encoder, and foot-kinematics noise covariances are calibrated offline.

The limitations are stated explicitly. The method uses only two modes and therefore cannot model intermediate slip or foot deformation without adding more hypotheses, introducing a cost-versus-distinguishability trade-off. It relies purely on proprioception, so large terrain geometries are not directly sensed. It also assumes that the rolling model is valid on most flat or uneven terrain; large local foot compliance may violate this assumption.

These constraints delimit the scope of proprioceptive scenario gating. The framework is not a general substitute for terrain perception, nor does it claim to identify all contact phenomena. Rather, it adapts online to slippage purely from IMU and joint-encoder data by combining rolling-aware ESKF models under multiple contact-mode noise assumptions. Within that scope, it is presented as a method for significantly improving legged-robot odometry in both simulation and real experiments.