⚡ electrical engineering

Terrain-Aware Convex Feasible Region Planning for Robust Wheeled-Legged Locomotion

This paper introduces a terrain-aware base-reference planner using convex feasible regions to generate long-horizon geometric guidance for wheeled-legged robots, significantly improving staircase traversal reliability compared to standard nonlinear model predictive control without increasing computational cost.

Original authors: Zhen Fu, Wei Zhang

Published 2026-09-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhen Fu, Wei Zhang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Robots that walk on wheels and legs are a promising solution for navigating the messy, unpredictable world humans inhabit. Unlike traditional wheeled vehicles that struggle on stairs or uneven ground, and unlike legged robots that can be slow and energy-hungry, these hybrid machines aim to roll efficiently on flat surfaces and step over obstacles when necessary. However, making a robot do this safely requires a brain that can see the ground ahead and plan a path that is physically possible for its body to follow. If the robot plans a movement that looks good on a computer screen but pushes its joints too far or places a wheel on a step it cannot reach, the machine will stumble or fall. The challenge lies in bridging the gap between a rough map of the terrain and the precise, split-second decisions a robot must make to keep moving without toppling over.

Researchers at the Southern University of Science and Technology have developed a new planning method specifically for these wheeled-legged robots to handle difficult terrain like staircases. They focused on a problem where standard robot controllers often fail: the disconnect between a robot's immediate movements and the geometry of the steps it is about to climb. Standard systems often rely on short-term guesses or simple rules to decide where the robot's body should go next. On a continuous, bumpy road, this works well. But on a staircase, where the ground is broken into distinct, separate levels, these short-term guesses can lead the robot to plan a path that requires its legs to stretch beyond their physical limits. The researchers found that by creating a "safe zone" for the robot's body based on the exact shape of the steps and the robot's own reach, they could guide the machine much more reliably.

The core of their solution is a system that translates the shape of the stairs into a series of invisible, three-dimensional boxes where the robot's body is allowed to be. First, the robot uses its cameras to build a digital map of the steps. It then identifies exactly where each wheel can safely land. Using these landing spots, the system calculates the range of positions the robot's main body can occupy while keeping its legs in a comfortable, reachable position. This is done by combining the shape of the step with the robot's leg capabilities to create a "convex feasible region," which is essentially a safe, solid volume of space. The system does this for every step the robot is about to take, creating a chain of safe zones that stretch far into the future.

Once these safe zones are defined, the robot uses a mathematical tool called a B-spline to draw a smooth path through them. Think of this as a flexible ruler that the computer bends to fit inside the safe zones, creating a gentle, continuous curve for the robot's body to follow. This curve acts as a long-term guide, telling the robot's lower-level controllers exactly where to be and how fast to move, ensuring that the robot never plans a step that would require its legs to stretch too far. This guide is calculated incredibly quickly, in less than two milliseconds, allowing the robot to update its plan in real-time as it moves.

The researchers tested this system on a four-legged robot with wheels, named W1, which has sixteen moving parts. They put the robot through a rigorous series of trials on a staircase with steps that were fifteen centimeters high and thirty centimeters deep. In one set of tests, the robot had to climb up and down the stairs continuously. The new system succeeded in every single attempt, completing two hundred trials without a single fall. In contrast, a standard robot controller using older methods succeeded in only about seventy-one percent of its attempts, failing over sixty times. The difference was even more dramatic in a second test where the robot had to climb the stairs, stop abruptly in the middle, hold its balance, and then continue. The new system stopped and resumed successfully one hundred times out of one hundred. The standard system managed to do this less than half the time, often losing its balance or failing to stop safely.

The researchers observed that when the standard system failed, it was usually because the robot planned a path that forced its wheels to cut across the edge of a step or required a leg to extend too far. The new system prevented these errors by strictly keeping the robot's body within the calculated safe zones. The success suggests that giving the robot a clear, long-term geometric guide is far more effective than trying to make the robot's short-term controller smarter or faster. While the system was tested on a specific robot and a specific type of staircase, the results indicate that this method of planning could significantly improve the reliability of robots navigating complex, stepped environments. The work does not claim to solve every possible terrain problem, and the researchers note that the system still relies on accurate maps and known step patterns, but for the specific challenge of climbing stairs, the new approach proved to be a decisive improvement over existing methods.

Technical Summary: Terrain-Aware Convex Feasible Region Planning for Robust Wheeled-Legged Locomotion

Problem Statement
Wheeled-legged robots offer a unique combination of rolling efficiency and legged terrain adaptability, making them suitable for complex environments like industrial facilities and disaster sites. However, operating on geometrically discontinuous terrain, such as steep staircases, presents significant challenges for standard perceptive Nonlinear Model Predictive Control (NMPC) frameworks. Existing approaches often rely on short prediction horizons (0.5–1.0 s) and heuristic base references that fail to encode kinematic feasibility relative to discrete footholds. This leads to kinematic inconsistencies, such as leg over-extension, when the terrain geometry is tight. Furthermore, for wheeled-legged systems, the continuous evolution of contact locations along the rolling direction adds complexity that point-foot models do not address. Current trajectory optimization methods either require prohibitive computation or use simplified models that ignore these specific kinematic limits.

Methodology
The authors propose a hierarchical control pipeline that integrates a Terrain-Aware Convex Feasible Region (CFR) Planner as a reference generation layer, decoupled from an existing kinodynamic NMPC and a Whole-Body Controller (WBC).

Terrain Perception and Foothold Extraction: The system processes depth data to generate a local 2.5D elevation map. Using a Raibert heuristic, it identifies nominal footholds and extracts bounded, collision-free convex polygons ( $S_{i,j}$ ) for each leg's scheduled contact phase.
Phase-wise CFR Construction: The core innovation maps these discrete foothold polygons into continuous base regions using conservative reachability approximations:
- Horizontal Reachability: The horizontal projection of the foothold is expanded via a Minkowski sum with a conservative 2D convex set of admissible foot displacements ( $\bar{W}_{kin}$ ) relative to the hip. The intersection of these expanded sets across all stance legs in a phase yields a horizontal convex polygon ( $Q^{xy}_j$ ).
- Vertical Reachability: Vertical bounds are calculated by intersecting the admissible height intervals derived from the vertical reach limits of all stance legs, resulting in a vertical interval ( $Q^z_j$ ).
- The complete phase-wise CFR is the Cartesian product $Q_j = Q^{xy}_j \times Q^z_j$ .
Convex B-Spline Optimization: The base trajectory is parameterized as a $k$ -th order 3D B-spline. The planner formulates a convex Quadratic Program (QP) where the control points are constrained to lie within the corresponding phase-wise CFRs. The objective function balances trajectory smoothness (minimizing the integral of the squared second derivative) with tracking of a nominal velocity reference. This QP is solved in real-time (typically < 1 ms).
Integration with NMPC and WBC: The resulting smooth 6-DoF base reference is fed into a low-rank contact-relaxation NMPC (running at 100 Hz), which computes ground reaction forces and end-effector trajectories over a short horizon. Finally, a full-body WBC (running at 500 Hz) maps these to joint torques, enforcing instantaneous no-slip constraints at the wheel-ground interface.

Key Contributions

Phase-Aware Convex Base-Region Construction: The paper introduces a method to map discrete foothold regions into per-phase base regions using Minkowski sums and set intersections. This transforms long-horizon geometric reasoning into linear constraints suitable for real-time optimization.
Real-Time B-Spline Trajectory Optimization: The authors demonstrate a convex QP formulation that generates long-horizon base references bounded by CFRs, executing in under 2 ms onboard. This provides geometric foresight to a short-horizon NMPC without extending the NMPC horizon itself.
Hardware Validation: The framework was validated on the W1 wheeled quadruped robot, showing significant improvements in success rates on staircases compared to a standard perceptive NMPC baseline.

Experimental Results
The proposed pipeline was tested on the W1 robot traversing a staircase with 15 cm step heights and 30 cm tread depths.

Stair Traversal: The proposed method achieved a 100% success rate (200/200 trials, including 100 ascents and 100 descents), compared to 71.36% (157/220 trials) for the baseline.
Mid-Stair Stopping: In trials requiring the robot to halt abruptly on the stairs and resume, the proposed method achieved 100% success (100/100), while the baseline succeeded in only 47.52% (48/101).
Performance: The B-spline planner solved in typically less than 1 ms (worst case < 2 ms). The NMPC ran at 100 Hz, and the WBC at 500 Hz, all meeting real-time deadlines.
Failure Analysis: Baseline failures were attributed to geometric inconsistencies where planned wheel trajectories intersected step edges, whereas the CFR-based planner successfully cleared these edges.

Significance and Scope
The paper claims that phase-wise geometric guidance significantly improves reliability on tested discrete terrain without requiring the computationally expensive extension of the NMPC horizon. The method effectively bridges the gap between discrete contact constraints and continuous base motion planning for wheeled-legged robots.

However, the authors maintain a modest scope regarding their claims:

The results are specific to the W1 platform, a single staircase geometry, and a specific externally supplied trot gait.
The CFR construction is a conservative, decoupled approximation; it does not provide a formal certificate of full-body kinematic or dynamic feasibility across phase boundaries.
The "fail-soft" mechanism (falling back to heuristic references when CFRs are empty) was not triggered during the reported trials and is not validated as a safety recovery controller.
Quantitative evaluation was limited to the staircase; demonstrations on ramps, gravel, and speed bumps were qualitative only.
The approach assumes known static terrain and accurate 2.5D elevation mapping, with sensitivity to reflective surfaces and poor depth measurements.

The work positions itself as a complementary layer to existing NMPC and WBC frameworks, offering interpretable, conservative reachability bounds rather than an end-to-end learned policy.

More like this