VOCALoco: Adaptive Locomotion for Robots

• VOCALoco is a modular framework that selects a specialized locomotion policy based on terrain viability and energy cost to ensure safe and efficient robot movement.
• It leverages egocentric heightfields and dual CNN predictors to estimate short-horizon safety and cost of transport over a fixed 1.5 m planning horizon.
• The framework enables smooth policy switching for tasks like stair ascent/descent and obstacle traversal, enhancing interpretability and energy efficiency.

VOCALoco, short for Viability-Optimized Cost-aware Adaptive Locomotion, is a modular skill-selection framework for legged robot locomotion that evaluates the viability and cost of transport (CoT) of a set of pre-trained locomotion policies on locally perceived terrain, then selects a policy that is both safe and energy-efficient (Wu et al., 28 Oct 2025). It is designed for settings such as stair ascent and descent, where conventional end-to-end deep reinforcement learning (DRL) policies can suffer from limited interpretability, limited flexibility, and no explicit notion of short-horizon safety. The framework operates on egocentric heightfields derived from perception, predicts per-skill traversability over a fixed planning horizon, and uses those predictions to gate or switch among specialized low-level controllers.

1. Problem formulation and conceptual basis

VOCALoco addresses a setting in which a quadrupedal robot has access to a repertoire of pre-trained locomotion policies, denoted

$\Pi = \{\pi_1, \pi_2, \dots, \pi_K\},$

where each policy is specialized to a terrain regime or locomotion behavior such as walking on flat or rough terrain, ascending stairs, descending stairs, climbing tall obstacles, or crossing gaps (Wu et al., 28 Oct 2025). Rather than training a single monolithic controller to map observations directly to torques across all terrain types, VOCALoco introduces a high-level decision layer that evaluates each available skill on the currently observed local terrain patch.

The motivating critique of end-to-end DRL is specific. A single policy is optimized statistically over a training distribution and does not possess an explicit notion that a region is unsafe for its current behavior. When facing out-of-distribution terrain, such a policy may crash rather than stop or adapt. The same monolithic structure also makes it difficult to interpret failures, difficult to add new behaviors without retraining a large network, and unable to reason explicitly about energy use at run time (Wu et al., 28 Oct 2025). VOCALoco therefore separates the problem into specialized low-level controllers and a compact high-level selector.

A common misconception is that VOCALoco replaces low-level locomotion control. It does not. The framework assumes access to pre-trained locomotion policies and only decides which policy to execute under the current terrain conditions. If no available policy is predicted to be safe, the framework can stop rather than attempt a traversal.

2. Viability, CoT, and the decision rule

The central quantities in VOCALoco are a viability predictor and a CoT predictor for each skill (Wu et al., 28 Oct 2025). Given a terrain patch $H$, the viability predictor

$V^k(H) \in [0,1]$

estimates the probability that executing skill $\pi_k$ is safe over a fixed horizon, while the CoT predictor

$C^k(H) \in \mathbb{R}^+$

predicts the cost of transport for the same skill and horizon.

Viability is defined operationally as the likelihood that a locomotion skill achieves the task of moving forward over a fixed horizon distance without any base collisions. In data collection, the robot is commanded to move straight ahead at a fixed velocity of $0.6$ m/s over a fixed horizon distance of $1.5$ m. A rollout is counted as a success if the robot reaches the target distance within a time limit of $4$ s and no base collision occurs. Repeating this $N$ times for a given skill and heightfield yields the empirical estimate

$$\hat V^k(H) = \frac{\# \text{success rollouts for skill }k}{N}.$$

The CoT model uses the standard definition

$H$0

where $H$1 is the vector of joint torques, $H$2 is the vector of joint velocities, $H$3 is the simulation time step, $H$4 is the robot’s mass, $H$5 is gravitational acceleration, and $H$6 is the distance traveled during the horizon, here $H$7 m (Wu et al., 28 Oct 2025). For CoT data collection, one rollout is performed per $H$8, the first $H$9 s of the rollout is ignored, and only non-crashing rollouts are retained.

The selection rule is a two-stage constrained choice. First, all skills with viability below a skill-dependent threshold are discarded. Then the skill with lowest predicted CoT among the safe skills is selected: $V^k(H) \in [0,1]$0 The paper reports typical thresholds

$V^k(H) \in [0,1]$1

Equivalently, the framework solves

$V^k(H) \in [0,1]$2

This formulation is important for interpretation. VOCALoco does not claim formal safety guarantees in the sense of verified control; rather, it uses learned short-horizon estimates of safe traversal probability and energy expenditure as explicit decision variables. That distinction is central to understanding the system’s scope.

3. Perception pipeline and terrain representation

VOCALoco uses egocentric heightfields derived from a 3D elevation map built from depth-camera perception (Wu et al., 28 Oct 2025). From the elevation map, the system extracts a rectangular terrain patch around the robot’s base. In the reported implementation, the patch spans 2 m forward, 1 m backward, 1 m wide, with 10 cm grid resolution, producing a heightfield

$V^k(H) \in [0,1]$3

The representation is normalized by subtracting the height at the center cell, so the heightfield encodes terrain shape relative to the robot’s local base altitude. This makes the predictor focus on local geometry rather than global elevation.

The paper also describes a specific treatment of occlusions. When ascending stairs, the robot cannot see beyond the top edge, leading to NaNs in the heightfield. These are handled with forward fill along columns: scanning from the back of the robot toward the front, NaNs are replaced with the last valid value behind them (Wu et al., 28 Oct 2025). The resulting dense, normalized heightfield is the sole input to all viability and CoT networks.

Perception noise is modeled during simulation data collection by perturbing each heightfield value with uniform noise in $V^k(H) \in [0,1]$4 m. There is no explicit probabilistic uncertainty model; robustness is instead learned through noise injection. This suggests a design that prioritizes simplicity of deployment over formal uncertainty propagation.

4. Network architecture, training, and control loop

For each skill $V^k(H) \in [0,1]$5, VOCALoco trains two independent CNNs with identical structure: one for viability and one for CoT (Wu et al., 28 Oct 2025). Each network takes $V^k(H) \in [0,1]$6 as input. The convolutional trunk consists of:

• Conv1: 4 channels, kernel $V^k(H) \in [0,1]$7, stride 1, padding 1
• Conv2: 8 channels, same kernel, stride, and padding
• MaxPool: kernel 2
• Conv3: 8 channels, same parameters

The MLP head contains FC1 with 128 units, FC2 with 128 units, and a scalar output. ReLU is used in all hidden layers. The viability CNN applies a sigmoid to map outputs to $V^k(H) \in [0,1]$8, whereas the CoT CNN outputs a raw scalar. Both networks are trained with mean squared error using stochastic gradient descent with learning rate $V^k(H) \in [0,1]$9 and momentum $\pi_k$0. The dataset size is 100k samples per CNN, and the paper reports about 12 minutes per CNN on an NVIDIA RTX 4090, with data collection taking about 4 hours for viability and 4 hours for CoT, and being parallelizable.

The low-level locomotion policies are trained with DRL rather than hand-designed control. The reported implementation uses PPO in Legged Gym, with Isaac Gym GPU-accelerated physics and the ANYmal-D quadruped (Wu et al., 28 Oct 2025). Three base skills are described:

• Walk policy: training terrain mix of 60% flat, 20% discrete, 20% rough terrain
• Ascend-stairs policy: 75% ascending stair terrains, 25% rough/discrete
• Descend-stairs policy: 75% descending stair terrains, 25% rough/discrete

At run time, skill selection is performed at 50 Hz. To avoid rapid switching caused by small fluctuations in predicted best skill, VOCALoco uses a sliding window filter over the last 10 predicted best skills. A switch is executed only when all entries in that window agree, which corresponds to a 0.2 s decision lag at 50 Hz. The paper characterizes this as providing fast responsiveness together with smooth transitions.

The fixed horizon of 1.5 m at nominal forward velocity $\pi_k$1 m/s is a deliberate compromise. The paper states that too short a horizon makes decisions noisy and switching too frequent, whereas too long a horizon makes predictions less accurate and less reactive to local changes.

5. Experimental results in simulation and on hardware

The experimental program includes simulation environments for low-level policy training, simulation environments for viability and CoT data collection, staircase transition courses for evaluation, and real-world deployment on ANYbotics ANYmal-D with zero-shot transfer from simulation (Wu et al., 28 Oct 2025). The baseline is a single end-to-end DRL locomotion policy from Legged Gym trained for rough terrain locomotion, using only proprioception and no explicit heightfield risk evaluation.

The reported findings separate naturally into predictor quality, behavior on stair courses, modular extension, and real-world performance. For viability prediction, predicted curves closely follow observed success rates in both ascending and descending staircase environments. As stair step height increases, viability decreases smoothly. For CoT, predicted CoT increases with step height, and the walk policy has lower CoT than stair-specialized skills on very small steps, consistent with lower swing heights and lower mechanical work.

On the staircase transition courses, VOCALoco exhibits three regimes. At small step heights it often selects the walk skill and traverses successfully. At intermediate step heights it detects that walk has become less viable while the stair-specialized policy remains viable, then switches to the appropriate stair policy and later back to walking on flat ground. At high step heights, when all policies have low viability, the system triggers Stop Activated, halting rather than attempting traversal (Wu et al., 28 Oct 2025).

The comparison with the Legged Gym baseline is explicit. As step height increases, the baseline’s success rate deteriorates sharply; it continues trying to track the desired velocity and often crashes or gets stuck. VOCALoco either achieves significantly higher success rates or correctly identifies the terrain as too difficult and stops rather than crashing. The paper further notes that at step height 22.5 cm, VOCALoco succeeds or crashes less often than the baseline.

The framework is also extended by adding two new skills: climb tall obstacles and cross gaps. These new policies use position-based target commands, and their viability and CoT networks are trained in the same manner. Crucially, the existing networks are not retrained. In a test obstacle course with a 0.5 m wall and a 0.5 m gap, VOCALoco walks on flat terrain, switches to the obstacle skill at the wall, switches to the gap skill at the gap, and then switches back to walking (Wu et al., 28 Oct 2025). This is the clearest demonstration of the framework’s modularity.

Real-world experiments use joystick commands specifying linear forward velocities between 0.5 and 0.6 m/s, with high-level selection and low-level control both running at 50 Hz. On a three-step staircase with steps of around 12 cm height, VOCALoco walks on flat ground, switches to ascend or descend near the staircase, and then switches back to walk after cresting or finishing the stairs. The paper reports that policy transitions are smooth, with no oscillation or instability observed.

6. Interpretation, limitations, and future directions

VOCALoco’s principal contribution is not a new locomotion policy class but a specific mode of integrating perception, risk prediction, and energy reasoning into a modular controller stack (Wu et al., 28 Oct 2025). Its interpretability comes from the fact that failures can be decomposed into at least two classes: the viability predictor may be wrong, or the low-level policy may be weak on a given terrain. This is structurally different from diagnosing a single monolithic DRL controller.

The framework also makes energy reasoning explicit. In many DRL systems, energy efficiency is shaped indirectly through reward design during training. VOCALoco instead predicts CoT at run time for each candidate skill and optimizes energy under a viability constraint. This supports behavior such as preferring the walk policy on small steps, where it is safe and cheaper, and switching to stair-specialized policies only when their higher viability justifies their higher cost.

Several limitations are stated or directly implied. VOCALoco depends on the availability and quality of pre-trained skills; it cannot synthesize a new behavior if all current skills are unviable. Its predictions are defined only over a fixed 1.5 m horizon, so longer-horizon planning is not modeled explicitly. The terrain representation is a heightfield, which does not capture overhanging structures and is oriented toward rigid ground geometries such as stairs, obstacles, and gaps. Perception uncertainty is handled through noise injection rather than explicit probabilistic inference. Finally, viability thresholds $\pi_k$2 require calibration to trade off conservativeness and aggressiveness (Wu et al., 28 Oct 2025).

The future directions identified in the paper include moving from discrete skill switching to policy mixing, extending the framework to more diverse terrain types including non-rigid or deformable surfaces, incorporating more sophisticated uncertainty handling and longer-horizon planning, and enabling online skill learning or adaptation. A plausible implication is that VOCALoco is best understood as a local decision layer that could be combined with broader navigation and planning systems rather than as a complete locomotion stack on its own.