Why legged robots stumble when switching between learned skills

Machines can master running or jumping in isolation, yet handing momentum safely between specialized motor routines mid-stride remains an open engineering problem.

A legged robot sprinting across an obstacle course looks remarkably competent until the moment it has to change what it is doing. Over the past few years, simulated trial and error has produced machines capable of navigating broken ground, executing parkour leaps, and striking balls in robotic soccer.¹ Yet when an agile biped needs to stop running and leap across a ditch, or transition from a steady walk into carrying a bulky container, the coordination often breaks down. The failure rarely stems from an inability to execute either individual movement; instead, the physical junction between two distinct motor skills proves surprisingly fragile.

This vulnerability matters because real environments demand continuous successions of varied physical tasks. When an automated machine operates in disaster response, construction, or domestic assistance, it cannot rely on a single, continuous routine. If every transition between motor routines poses a severe hazard of falling, machines with legs remain confined to isolated demonstrations or heavily scripted routines. Resolving how independent motor skills hand off physical momentum could turn separate mechanical tricks into an adaptable repertoire.

To understand why physical handoffs are difficult, consider the sequence of physical conditions governing any legged gait. A running humanoid robot maintains forward balance by timing footfalls to counteract gravity, relying on dynamic momentum. Its joints, motors, and sensors continuously adjust torques to preserve that forward trajectory. If a jump becomes necessary, the limbs must generate a rapid upward impulse, which requires positioning the feet, bending the knees, and absorbing incoming kinetic energy. When a machine switches routines mid-stride, the incoming jumping controller suddenly inherits an arbitrary physical state: one leg may be trailing in mid-air while the torso leans forward at an awkward tilt. Without an explicit way to reconcile that sudden mismatch, the incoming controller commands torques ill-suited to the current posture, causing the machine to lose balance.

To bypass this dilemma, roboticists typically resort to a clumsy compromise: bringing the machine to a complete standstill before starting the next movement. Halting settles the body into a stable, predictable stance where the limbs rest symmetrically, eliminating unexpected velocities. The incoming routine can then activate safely from a resting position. This stop-and-go pattern resolves the mathematical problem of safety, but it discards all accumulated forward speed, making dynamic movements such as a leap taken during a sprint impossible.¹

Why do separate robot controllers fail to connect smoothly?

Separate robot controllers fail to connect smoothly because each policy is trained in isolation to handle only its own specific operational regime.¹ In a conceptual paper posted to the arXiv preprint server, Daniel Gigliotti, Flavio Maiorana, Fabio Patrizi, and Luca Iocchi, computer engineers at Sapienza University of Rome, argued that skill composition should be investigated as a primary research challenge rather than treated as an incidental programming task. The researchers pointed out that training a single, monolithic artificial neural network using deep reinforcement learning to govern every conceivable movement creates severe mathematical hurdles.¹ Such end-to-end policies require complex composite reward functions balancing safety, forward progress, and mechanical stability all at once.¹ Over extended timelines, determining which specific motor adjustment was responsible for a later fall becomes mathematically intractable.¹

Training specialized sub-policies, by contrast, offers significant practical advantages. A sub-policy, which is an independent neural network controller trained to master a single discrete behavior, learns rapidly because its objective remains focused.¹ Engineers can evaluate and validate that individual skill in simulation. When roboticists keep these sub-policies frozen, preserving their internal weights without modification, the library of movements can expand simply by adding new skills.¹ However, when an external instruction demands an immediate transition between two frozen skills, the physical state left behind by the outgoing movement often falls outside the acceptable starting parameters of the incoming one.¹

Engineers have previously attempted to resolve this junction problem by adjusting the policies so that they accommodate each other.¹ Some teams fine-tune the neural networks so that the end of one skill matches the beginning of another.¹ Others train additive correction networks on top of frozen controllers.¹ As Gigliotti and colleagues noted in their review, those methods achieve smoother transitions by sacrificing modularity.¹ Once individual behaviors are altered to interlock with specific companions, they can no longer be updated or reused independently across different tasks without retraining the surrounding software.

How can machines link motor behaviors without stopping?

Machines can link motor behaviors without stopping by deploying dedicated connection mechanisms that coordinate the handover of physical control between independent sub-policies.¹ The researchers from Sapienza University of Rome identified two distinct mathematical arrangements for combining specialized controllers: blending and bridging. Each approach delegates execution to a decision layer, but they handle physical transitions through fundamentally different strategies.

Blending combines concurrently active controllers by calculating a continuous weighting across several outputs.¹ A decision layer assigns proportional weights to different sub-policies, merging their simultaneous motor commands into a single physical action.¹ This strategy fits tasks involving the simultaneous coordination of different bodily regions, such as locomotion combined with object manipulation, where one controller directs the legs while another guides the arms.¹ Because the weights shift along a continuous spectrum, the merged command can absorb changing task demands. Yet blending encounters serious obstacles when two sub-policies command the same physical joints, creating conflicting mechanical forces that neither controller would produce alone.^contributed

Co-author Luca explained that evaluating these interactions remains an unresolved technical hurdle. Luca said in response to questions from Primary that "How to measure the performance of the different blending techniques is an open research question."^contributed Luca added that "Empirical observations show macro failures (e.g., humanoid robot falling down) in circumstances (i.e., combinations of merged sub-policies and blending techniques) that are difficult to formally identify."^contributed Luca also observed that "blending in quadruped robots is generally more robust than when used on humanoid robots, but at this moment we are not able to scientifically answer the question about why."^contributed

Bridging provides an alternative mechanism by executing an abrupt handoff between sequential controllers.¹ In this framework, the decision layer selects a single active skill that retains sole control until an external signal selects a replacement. To manage the sudden shift, the system inserts a short-lived connective routine called a bridge.¹ When an incoming skill is summoned, the bridge takes command instantly from whatever unstable posture the outgoing skill created.¹ Its sole objective is to guide the robot into a physical condition from which the incoming controller can operate safely. Once the machine enters a viable starting state, the bridge hands over execution and deactivates.¹

Traditional robotic frameworks assume that an outgoing routine runs until reaching a planned termination point.¹ Under that older model, the exiting controller deliberately steers the robot toward a state where the next controller can safely begin. Gigliotti and colleagues emphasized that real-world operations discard this guarantee because an external supervisor or unpredictable obstacle can demand a behavioral change at an arbitrary millisecond.¹ An effective bridge must therefore handle unexpected interruptions, taking control when limbs and momentum are mid-motion.

What makes choosing an entry state so challenging?

Choosing an entry state is challenging because the connective controller must look ahead and select an advantageous physical configuration rather than merely waiting for a tolerable moment to yield control.¹ Gigliotti and colleagues observed that existing transition methods focus exclusively on the timing of a switch.¹ Those systems hold the transition until the machine drifts into an acceptable posture, but they do not actively steer the body toward a target state that optimizes the incoming movement. In dynamic actions like running into a jump, an incoming leaping routine can start from a broad family of physical positions, yet some configurations exploit existing forward momentum while others waste kinetic energy.¹

Actively selecting an optimal landing posture requires solving two open technical problems: comparison and anticipation. First, a controller needs a reliable method to measure whether one body configuration serves the next behavior better than another. Standard geometric measurements of joint angles prove inadequate, because two physical states that appear nearly identical can diverge rapidly into completely different physical trajectories.¹ Instead, states must be evaluated based on whether applying the incoming behavior leads to comparable dynamical outcomes.

Luca pointed out that physical divergence arises from unmodelled environmental forces and numerical noise during simulation. Luca told Primary that "This divergence is mainly due to dynamics and sensor noise."^contributed Luca explained that "Dynamics introduces non-stationarity (breaking the assumptions made in al RL algorithms). Sensor noise accumulates over time, but also over space (e.g., joint position errors accumulated toward the end-effector position)."^contributed To address these discrepancies during early development, the group relied on privileged simulation data, such as exact center-of-mass positions and foot velocities. Luca noted that transferring these solutions to physical robots creates another research challenge, stating that "how to fill sim-to-real gap when training in simulation uses privileged information that are not available in the real deployment" must be addressed with dedicated transfer methods.^contributed

Second, the system needs anticipation. To choose the best posture for the handover, the bridge must know what the incoming skill will do once it takes charge. Gigliotti and colleagues proposed that each specialized skill could carry a compact descriptor, calculated once during training, that summarizes how that skill will behave when launched from various postures.¹ A bridge controller could then consult those precomputed behavioral descriptors to select the most favorable physical state, transforming an unpredictable trial-and-error handoff into a deliberate, calculated transition.

Luca indicated that the formulation of compact descriptors remains an active area of investigation. Luca said in response to questions from Primary that "This idea is still under development. The general concept is computing policies that are conditioned to the expected performance in the future in another task."^contributed Luca explained that their current approach defines viable starting configurations by collecting rollouts of sub-policies initialized from randomized starting states, emphasizing that the underlying mathematics still require deeper study.

What does this framework leave unresolved?

This conceptual paper represents an architectural framework and literature analysis rather than an empirical demonstration on physical hardware. The researchers reported preliminary investigations across two simulated testbeds, including a blended kicking behavior and bridged movements, but the full engineering architecture remains under development.¹ The publication appeared as a workshop paper for the Artificial Intelligence and Robotics workshop at AIxIA 2026, posted to the arXiv preprint server without formal peer review.¹ As a result, its central arguments remain theoretical propositions about software design rather than verified physical benchmarks.

Luca confirmed that broader software access and rigorous experimental trials are still underway. Luca told Primary that "Code development is also in progress and its release requires additional work to make it ready for publication."^contributed Luca noted that the research group is conducting comprehensive comparisons against existing state-of-the-art methods while refining the codebase.

The paper also leaves open the broader question of how bridging controllers should scale across an extensive library of movements. Designing an individual bridge for every possible pair of skills would create combinatorial complexity, requiring dozens of specialized transition networks as a repertoire expands. The authors acknowledged that finding whether a single generalized bridge can manage handoffs between arbitrary skills, while maintaining balance and shedding excess momentum, remains an unanswered empirical challenge.¹ Luca remarked that "we do not know yet how to formally characterize bridging properties, how to measure its performance, how to isolate failures, etc."^contributed Luca stressed that while empirical tests reveal where transitions fail, "we also need more theoretical studies to understand the exact mathematical properties that govern its behavior."^contributed

Furthermore, the paper did not establish rigorous statistical metrics for evaluating transitions independently from overall task success.¹ Luca stated that while an ideal theoretical benchmark would compare a handover trajectory against an expensive end-to-end controller trained on the combined task, practical benchmarks will likely rely on measuring differences between the transition policy and the consecutive sub-policies.

How could modular skills change robotic autonomy?

Establishing dependable bridges between motor skills could fundamentally alter how high-level software governs automated machines. In recent robotics research, engineers have explored delegating behavioral choices to large vision-language models.¹ While those neural networks interpret complex environments effectively, they operate as opaque black boxes whose subsequent physical choices cannot be inspected or verified in advance.

If physical execution can be handed off safely at any moment through verifiable bridges, high-level guidance can instead be assigned to transparent computational tools. Symbolic planners, automata, and formal logic controllers could determine what the machine should do next according to stated safety rules.¹ Because the bridge guarantees that the physical body can safely absorb a command issued mid-stride, the high-level planner does not need to compute complex joint torques.¹

Luca emphasized that this approach merges abstract reasoning and physical reflexes into an integrated framework. Luca told Primary that "Overall this research leads toward the direction of neuro-symbolic planing and acting, where high-level logic and low-level skills are processed as a unified process, moving forward from the clear separation between high-level sub-tasks and low-level continuous actions."^contributed In that system, the quality of a physical bridge directly shapes high-level logic, allowing an autonomous planner to select alternative plans if a particular motor transition is deemed unreliable. By separating high-level strategic reasoning from low-level mechanical reflexes, robotic systems could become both physically agile and mathematically auditable.¹

What this rests on

50 sentences trace to 1 source and 1 contributor.

All 2Documents 1Records 0Reporting 0People 1

1. 1 DG Daniel Gigliotti, Flavio Maiorana, Fabio Patrizi, Luca Iocchi Paper · arxiv.org · 15 Sep 2026 Skill Composition for Legged Robot Reinforcement Learning Preprint · may not have been peer reviewed See the source
2. 2 L Luca Contributor · interview 28 Sep 2026 · in the article since 2 Oct 2026 Contribution — Luca 18 statements added to this article

Article history

1. Published 2 Oct 2026, 17:53
   Assembled by the Primary desk from 1 source · 1 contributor · 50 cited sentences
2. L Luca 18 statements added 28 Sep 2026 5 days before publication

   1. Go to the sentence: Yet blending encounters serious obstacles when two sub-policies command the same physical joints, creating conflicting mechanical forces that neither controller would produce alone.↑
   2. Go to the sentence: Luca said in response to questions from Primary that "How to measure the performance of the different blending techniques is an open research question."↑
   3. Go to the sentence: Luca added that "Empirical observations show macro failures (e.g., humanoid robot falling down) in circumstances (i.e., combinations of merged sub-policies and blending techniques) that are difficult to formally identify."↑
   4. Go to the sentence: Luca also observed that "blending in quadruped robots is generally more robust than when used on humanoid robots, but at this moment we are not able to scientifically answer the question about why."↑
   5. Go to the sentence: Standard geometric measurements of joint angles prove inadequate, because two physical states that appear nearly identical can diverge rapidly into completely different physical trajectories.↑
   6. Go to the sentence: Luca told Primary that "This divergence is mainly due to dynamics and sensor noise."↑
   7. Go to the sentence: Luca explained that "Dynamics introduces non-stationarity (breaking the assumptions made in al RL algorithms). Sensor noise accumulates over time, but also over space (e.g., joint position errors accumulated toward the end-effector position)."↑
   8. Go to the sentence: Luca noted that transferring these solutions to physical robots creates another research challenge, stating that "how to fill sim-to-real gap when training in simulation uses privileged information that are not available in the real deployment" must be addressed with dedicated transfer methods.↑
   9. Go to the sentence: Gigliotti and colleagues proposed that each specialized skill could carry a compact descriptor, calculated once during training, that summarizes how that skill will behave when launched from various postures.↑
   10. Go to the sentence: Luca said in response to questions from Primary that "This idea is still under development. The general concept is computing policies that are conditioned to the expected performance in the future in another task."↑
   11. Go to the sentence: The researchers reported preliminary investigations across two simulated testbeds, including a blended kicking behavior and bridged movements, but the full engineering architecture remains under development.↑
   12. Go to the sentence: Luca told Primary that "Code development is also in progress and its release requires additional work to make it ready for publication."↑
   13. Go to the sentence: The authors acknowledged that finding whether a single generalized bridge can manage handoffs between arbitrary skills, while maintaining balance and shedding excess momentum, remains an unanswered empirical challenge.↑
   14. Go to the sentence: Luca remarked that "we do not know yet how to formally characterize bridging properties, how to measure its performance, how to isolate failures, etc."↑
   15. Go to the sentence: Luca stressed that while empirical tests reveal where transitions fail, "we also need more theoretical studies to understand the exact mathematical properties that govern its behavior."↑
   16. Go to the sentence: Furthermore, the paper did not establish rigorous statistical metrics for evaluating transitions independently from overall task success.↑
   17. Go to the sentence: Because the bridge guarantees that the physical body can safely absorb a command issued mid-stride, the high-level planner does not need to compute complex joint torques.↑
   18. Go to the sentence: Luca told Primary that "Overall this research leads toward the direction of neuro-symbolic planing and acting, where high-level logic and low-level skills are processed as a unified process, moving forward from the clear separation between high-level sub-tasks and low-level continuous actions."↑