MUJICA: Multi-skill Unified Joint Integration of Control Architecture for Wheeled-Legged Robots

Abstract: Wheeled-legged robots hold promise for traversing complex terrains and offer superior mobility compared to legged robots. However, wheeled-legged robots must effectively balance both wheeled driving and legged control. Furthermore, due to noisy proprioceptive sensing and real-world motor constraints, realizing robust and adaptive locomotion at peak performance of motors remains challenging. We propose the Multi-skill Unified Joint Integration of Control Architecture (MUJICA), a unified, fully proprioceptive control framework for wheeled-legged robots that integrates diverse low-level skills-including omnidirectional moving, high platform climbing, and fall recovery-within a single policy. All skills, distinguished by unique indicator variables, are trained jointly with accurate DC-motor constraint modeling. Additionally, a high-level skill selector is learned to dynamically choose the optimal skill based solely on proprioceptions, enabling adaptive responses to the surrounding environment. Therefore, MUJICA enhances sim-to-real robustness and enables seamless transitions across diverse locomotion modes, facilitating autonomous adjustment to the environment. We validate our framework in both simulation and real-world experiments on the Unitree Go2-W robot, demonstrating significant improvements in adaptability and task success in unstructured environments.

• The paper introduces MUJICA, a control framework for simulating and deploying wheeled-legged robots, jointly learning multiple heterogeneous skills through one policy with DC-motors limits
• MUJICA utilizes a multi-target proprioceptive estimator, DC-motor constraints, succeeding in simulating to real-world deployment with over 90% of the reducer overall violations on thighs below 3.5%, real motor faults successfully prevented by accurately mapping simulation failures to the real world
• Tests on real hardware Go2-W demonstrate 95% success rate in transitioning between recovery, omnidirectional motion, and platform climbing on diverse and challenging environments (20cm steps, 80cm and 1-meter gaps).

MUJICA is a unified, fully proprioceptive control framework for wheeled-legged robots that jointly learns three heterogeneous skills—omnidirectional locomotion, high platform climbing, and fall recovery—within a single policy, augmented with hard DC-motor constraints and a learned high-level skill selector. The work targets the Unitree Go2-W and demonstrates zero-shot sim-to-real deployment, including climbing platforms up to 1 m in height (2605.13058).

Motivation and problem formulation

The authors identify three shortcomings of existing legged and wheeled-legged controllers. First, multi-skill blind policies typically cover only similar or low-difficulty skills (e.g., stairs and slopes). Second, handling dissimilar skills with distinct dynamics generally requires manual switching or teacher-student distillation / mixture-of-experts pipelines, which cannot resolve conflicts between low-level behaviors and add network complexity. Third, most constrained-RL approaches impose simplified actuation limits—clipping torque to a constant maximum—that ignore the velocity- and position-dependent torque envelope of DC motors, limiting both performance exploitation and sim-to-real safety.

The problem is cast as a Constrained POMDP: maximize discounted return subject to expected constraint costs below thresholds $\delta_i$. The observation space contains only proprioceptions (base angular velocity and gravity vector, commands, joint positions/velocities, previous actions, skill indicator), while critics receive privileged observations including base linear velocity, contact information, and wheel-ground heights. Leg joints receive position-offset targets; wheels receive velocity targets through PD control.

Method

State estimator. A GRU-based online encoder processes the past $H=6$ observations and predicts four quantities from proprioception alone: base linear velocity, per-segment collision probabilities, wheel-ground distances, and a latent state vector supervised by a reference encoder via a SwAV contrastive loss. The paper argues that each target carries distinct task-relevant information: velocity estimation distinguishes stable from unstable regimes, wheel-ground distance reflects terrain roughness and suppresses unnecessary leg lifting on flat ground, and collision probability enables detecting ledges with head contact and triggering recovery.

Asymmetric actor-critic with P3O. The actor consumes the estimator embedding plus the current observation; both reward and constraint critics consume privileged observations. Optimization uses P3O, a penalty-based transformation of constrained RL into clipped PPO-style surrogate objectives for reward and for each constraint advantage. Constraints include a DC-motor torque constraint over all 16 actuators and thigh/calf collision constraints (the latter encouraging wheel rather than leg contact). The DC-motor model enforces a velocity-dependent torque limit—constant at low speed, decreasing linearly at high speed—with an additional cosine dependence on joint position for calf joints, using values from Unitree's motor manual.

Skill indicator and selector. Each skill receives a unique indicator variable $\zeta_t$ appended to the observation, disentangling task-specific behavior along this dimension. Training proceeds in two stages: S1 trains the low-level skills jointly with curriculum learning across a 33×20 grid terrain (stairs, slopes, discretized bumps, rough terrain, pits) with 4,096 parallel IsaacLab environments; S2 freezes the skills and trains a high-level selector—an asymmetric actor-critic over the last $H$ observations—to output a distribution over skills using unified velocity-tracking rewards. Domain randomization covers friction, mass bias, external forces, pushes, and motor gain multipliers.

Simulation results

All methods were trained under identical terrain curricula and rewards and evaluated on five tasks at ten difficulty levels with gaps doubled relative to training, over 4 seeds. MUJICA outperforms DreamWaQ + P3O, vanilla PPO, and ablations removing any single estimator target, with the gap widening at higher difficulties. Removing velocity estimation is particularly damaging for non-recovery tasks. Notably, even vanilla PPO solves slope traversal—a consequence of the wheeled morphology's mobility—but fails entirely at fall recovery due to the absence of state estimation.

The DC-motor constraint ablation provides the strongest quantitative result: without it, over 90% of thigh-joint torque–velocity samples lie outside the feasible motor region, with persistent oscillations and rear-thigh joint-limit impacts that trigger real motor faults during platform climbing. With the constraint enforced, violations drop below 3.5%. This directly links the constraint model to hardware survivability, not merely simulation cleanliness.

Skill selector evaluation

On sequential episodes chaining stair recovery → stair climbing → platform climbing at middle difficulty, the full method achieves 95%, 95%, and 91% stage-wise success rates, versus 72%, 72%, and 38% for a no-indicator variant trained with separate per-task rewards. A no-indicator variant with unified rewards frequently collapses during training altogether. These results indicate that reward engineering alone does not disentangle heterogeneous objectives; the explicit indicator is what stabilizes joint training.

Real-world experiments

Policies deploy zero-shot on Go2-W with all computation on a Jetson Orin NX at 50 Hz. Reported demonstrations include recovery from an upside-down state on 30° stairs, ascent of irregular 20-cm stairs, clearing an 80-cm outdoor platform and a 1-m indoor platform—the latter described as beyond prior wheeled-legged work, approaching mechanical and actuation limits via hind-leg energy storage and foreleg hooking. Across all tasks the motor protection mechanism never triggered, supporting the claim that accurate constraint modeling permits exploiting near-limit actuation safely. In continuous missions, the selector autonomously transitions recovery → omnidirectional motion → platform climbing based solely on proprioception.

Limitations and open questions

Several caveats are evident. All quantitative comparisons are confined to a single platform (Go2-W); generalization across morphologies is untested despite the framework's claimed extensibility. The evaluation terrains, while challenging, are structured (stairs, pits, platforms); performance on fully unstructured natural terrain is addressed only qualitatively and left as stated future work. The skill set is fixed at three discrete skills with hand-assigned indicators, so the approach depends on a predefined skill taxonomy—how new skills would be added without retraining S1/S2 remains open. Finally, the selector is trained on frozen skills under velocity-tracking rewards; whether hierarchical composition degrades when skill objectives conflict at boundaries (e.g., mid-climb command changes) is not characterized.

Conclusion

MUJICA demonstrates that a single blind policy can span substantially different wheeled-legged skills when equipped with (i) a multi-target proprioceptive estimator covering velocity, collisions, and wheel-ground distance, (ii) hard, physically accurate DC-motor constraints—which reduce simulated constraint violations from over 90% to below 3.5% and prevent real motor faults—and (iii) an indicator-conditioned architecture with a learned selector achieving up to 95% success on sequential multi-skill missions. The central open question is whether the indicator-based joint training paradigm scales to larger, more finely grained skill repertoires and less structured environments without sacrificing the stability that makes it effective here.