Implicit Behavior Coordination from Sub-Task Demonstrations by Exploiting Overlap-Induced Multimodality

Long-horizon robotic rearrangement tasks are often treated as skill sequencing problems, requiring predefined skills, skill labels, or boundaries, and task-specific switching logic. Although effective, such explicit skill abstractions can become difficult to scale as the number of behaviors and the task horizon increase. We instead formulate rearrangement as implicit-behavior coordination from unlabeled sub-task demonstrations, where skill-like behaviors are learned directly from mixed behavior data and coordinated through value-guided action selection. Experiments in Habitat rearrangement tasks support this formulation in three ways. First, our method outperforms task-specific imitation baselines on more complex rearrangement tasks and approaches an oracle-planner baseline with behaviorcloned skills, while using no oracle task plan or skill-labeled full-task demonstrations. Second, ablations show that reliable critic-guided candidate selection is essential for coordinating multi-modal behaviors. Third, scaling experiments show that the method handles larger behavior repertoires and maintains stronger performance than task-specific imitation baselines as chained targets extend the horizon. These results suggest that explicit skill abstraction is not a prerequisite for long-horizon rearrangement, and that implicit-behavior coordination offers a promising data-driven alternative to explicit skill-based pipelines.

We consider long-horizon robotic rearrangement in a partially observed setting. The initial object position and desired target position are provided, but the object may not be directly accessible, e.g., enclosed inside a receptacle, and no environment map or obstacle locations are given. Task objective is defined by placing the object at the target location. At each time step, the robot observes the relative positions of the object and target, arm joint positions, a binary gripper-holding indicator, and depth images from sensors mounted on the robot’s head and end effector. The action consists of base linear and angular velocities, incremental arm joint commands, and a binary gripper command. Unlike prior work that assumes complete task trajectories or explicit skills and coordination rules, we consider a more restrictive learning setting where the agent is given only unlabeled demonstrations of individual sub-tasks, such as navigation, receptacle opening, picking, and placing, together with the task objective. Each demonstration is a sequence of observations and actions, without subtask labels, skill boundaries, or full-task ordering annotations. The goal is to learn a policy that maps current observations to robot actions and adaptively coordinates the behaviors present in the demonstration data according to the task objective. Because the training data contains only sub-task demonstrations rather than complete task executions, the main challenge is to learn not only the individual behaviors but also how to coordinate them toward the sparse rearrangement objective.

We propose a framework that combines a multi-modal generative policy with a learned critic for adaptive coordination, as illustrated in Fig. 1. Unlabeled demonstrations from different subtasks contain overlapping observations with different corresponding actions. To model this multimodality, we use a conditional Flow Matching policy [17], which generates multiple candidate actions corresponding to different plausible behaviors. The critic evaluates these candidates under the task objective and selects the one with the highest value. Repeating this process allows the robot to coordinate different behaviors without complete task trajectories or hand-crafted coordination rules.

The Multi-Modal Flow Matching Policy is trained on mixed sub-task demonstrations, where each demonstration trajectory is converted into context–chunk samples (sn, an), where sn denotes a context of the previous C observations and an the corresponding chunk of H future actions. The policy predicts the velocity field from xτ and τ, conditioned on tokens encoded from visual encoders for depth images and a linear projection for non-visual observations. At inference, K action chunks are initialized from Gaussian noise and denoised with the context-conditioned policy. The generated chunks are then evaluated by the critic, and the chunk with the highest predicted value is selected for execution.

The Critic Network estimates the expected return, that is, the discounted sum of future rewards, of executing an action chunk an at observation context sn with respect to the task objective. To propagate reward information across different demonstrations, we adopt a in-sample planning method modified for our context–chunk settings. The return target of each context–chunk sample is updated as:

(i−1)

(i−1)

(i)

Rn(i) = rn + γ max Rn+, Vn+

,Vn+:= Ea∼π(·sn+) Qϕ (sn+, a),

and the critic is trained by minimizing the squared error between the predicted value Qϕ (sn, an) and the updated return target Rn. This update rule propagates high return values in two ways: first, reward information is propagated backward through the same demonstration trajectory via Rn+, and second, at overlapped contexts, the multimodal policy can generate high-value actions aligned with other demonstrations, producing a large next context value Vn+.

We also modified the computation of the next context value Vn+ by using an uncertainty-weighted average over the K candidate action chunks generated by the Flow Matching policy at the next observation context s+n. The weight of each candidate and the value of the next context are computed as:

(σ 2)−1,

wi = PK i

2 −1, j=1 (σj)

Vn+ =

K

X

wi Qϕ (s+n, ai)i=1. This reduces the influence of anomalous high-value action chunks that are not supported by the demonstrations and leads to more reliable return propagation.

The experiments evaluate whether implicit-behavior coordination can serve as a scalable alternative to explicit skill-based coordination for long-horizon rearrangement, focusing on three questions: first, can the proposed method coordinate behaviors from unlabeled sub-task demonstrations and outperform learning-based baselines that rely on skilllabeled complete-task demonstrations? second, does the method remain effective as the behavior repertoire grows, without adding skill labels or task-specific coordination logic? third, can the same framework scale to longer horizons with chained rearrangement objectives that require repeated behavior coordination? The results show strong coordination using unlabeled sub-task demonstrations and that implicit behaviors learned from unlabeled sub-task demonstrations can be coordinated to solve complex tasks, scale to larger behavior repertoires, and remain effective over longer horizons. For instance, in the Task-Horizon Scaling with Chained Targets experiment on Nav-Pick-Nav-Place tasks, our method maintains 35% success at five targets without oracle task plans or explicit skill decomposition or skill-labeled full-task demonstrations, whereas Skill Transformer degrades sharply from 66% for one target to 1% for five targets. This suggests that value-guided implicit-behavior coordination is more robust to repeated behavior transitions and accumulated execution errors than task-specific full-task imitation.

In conclusion, we presented implicit-behavior coordination, a formulation for learning long-horizon behavior coordination from unlabeled sub-task demonstrations. Our experiments suggest that long-horizon rearrangement does not necessarily require an explicit skill sequencing interface: trained only on mixed unlabeled sub-task demonstrations, the proposed framework allows efficient behavior transitions to emerge through value-guided selection over generated action candidates. Implicit-behavior coordination offers a middle ground: it avoids oracle task plans and skill-labeled full-task data while maintaining stronger performance over larger behavior repertoires and longer chained-target horizons. Future progress will require broader behavior datasets, stronger value grounding under sparse objectives, and robust object and target grounding for real-world rearrangement.

The paper also includes a Real-world Validation on a UR3e tabletop platform to examine whether the proposed framework can be instantiated beyond the simulated mobile-manipulation benchmark, providing preliminary evidence that the learned policy can execute the behavior sequence under real-world sensing and control noise. The limitation noted is that our method depends on sufficient coverage and overlap among sub-task demonstrations; sparse or poorly connected data can limit value propagation and behavior stitching. Furthermore, in the real-world setup, we do not have information about the target position and it is not included in the observations as in simulation. The target position can be only inferred through the sparse reward. That is why placing the object at different position, requires training a new critic with the new sparse reward. However, there is no need to collect more demonstrations or retrain the Flow Matching policy.

The key finding regarding coordination is that the critic must both coordinate implicit behaviors and propagate sparse reward across overlapping demonstrations. The uncertainty-weighted critic achieves the best performance, demonstrating the importance of value-guided action selection with reliable return propagation. This result is summarized by Table 2: Multiple candidates + uncertainty-weighted critic (Ours) 68.5±2.3.

The paper also highlights that implicit coordination requires reliable value-guided selection. This is demonstrated by Table 2, where the uncertainty-weighted critic achieves the best performance, improving success to 68.5% compared to single candidate or random selection which achieved only 61.5%. The scaling experiments further show that the full-repertoire agent achieves performance close to the corresponding taskspecific agents across all tasks, suggesting that adding additional unlabeled behavior demonstrations does not substantially degrade coordination performance."

The paper concludes that "implicit-behavior coordination offers a middle ground: it avoids oracle task plans and skill-labeled full-task data while maintaining stronger performance over larger behavior repertoires and longer chained-target horizons. The final conclusion is: Implicit-behavior coordination offers a promising data-driven alternative to explicit skill-based pipelines."

The paper's title, as provided in the prompt, is Implicit Behavior Coordination from Unlabeled Sub-Task Demonstrations by Exploiting Overlap-Induced Multimodality. This aligns with the core methodology described. The paper's abstract states: "We instead formulate rearrangement as implicit-behavior coordination from unlabeled sub-task demonstrations, where skill-like behaviors are learned directly from mixed behavior data and coordinated through value-guided action selection."

The relevant contributions are: "(i) We introduce an implicit-behavior coordination formulation for learning long-horizon behavior coordination from unlabeled sub-task demonstrations, replacing explicit skill sequencing with coordination over learned action candidates. (ii) We instantiate this formulation with a generative behavior model and value-guided action selection, as illustrated in Fig. 1, enabling coordination without complete-task trajectories, skill labels, or oracle task plans. (iii) We apply the framework to long-horizon robotic rearrangement as a testbed for studying whether explicit skill definitions and labels are necessary for behavior coordination."

The paper's title is Implicit-Behavior Coordination from Unlabeled Sub-Task Demonstrations by Exploiting Overlap-Induced Multimodality. This aligns with the core methodology described. The abstract states: "We instead formulate rearrangement as implicit-behavior coordination from unlabeled sub-task demonstrations, where skill-like behaviors are learned directly from mixed behavior data and coordinated through value-guided action selection."

The paper's title is Implicit Behavior Coordination from Unlabeled Sub-Task Demonstrations by Exploiting Overlap-Induced Multimodality. This aligns with the core methodology described. The abstract states: "We instead formulate rearrangement as implicit-behavior coordination from unlabeled sub-task demonstrations, where skill-like behaviors are learned directly from mixed behavior data and coordinated through value-guided action selection."

The paper's title is "Implicit Behavior Coordination from Unlabeled Sub-Task Demonstrations by Exploiting Overlap-Induced Mult

Here are the specific improvements that can be made to AI systems based on this research, and what those improved systems can achieve:

) Improved System Capabilities: Implicit-Behavior Coordination Framework

The core improvement is shifting from explicit skill sequencing/labeling to a data-driven, implicit coordination mechanism. The resulting AI system is characterized by its ability to learn complex, sequential behaviors directly from mixed, unlabeled demonstrations without requiring manual skill definitions or task plans.

Specific improvements and capabilities include:

A unified policy capable of generating multiple plausible action trajectories (candidate chunks) simultaneously for any given state, effectively modeling the multimodal nature of sub-task transitions (e.g., simultaneously considering navigation and picking actions).

Adaptive, value-guided action selection enabled by a learned critic that evaluates these generated candidates against the overall task objective (sparse reward), allowing the robot to dynamically choose the most promising behavior at each step, even when encountering novel or overlapping demonstrations.

Robust coordination across diverse and growing behavior repertoires (e.g., handling more types of objects or manipulation techniques) without needing explicit skill labels or re-engineering the task planner/coordinator layer for every new scenario.

) Specific Improvements in Application: Long-Horizon Robotic Rearrangement

This framework directly improves AI systems designed for complex, multi-step physical tasks that were previously bottlenecked by rigid skill hierarchies.

Handling highly complex, multi-modal rearrangement tasks (e.g., Nav-Open-Pick-Nav-Place) where success depends on the correct interleaving of distinct behaviors (navigation, grasping, opening a drawer). The system can now effectively stitch these sub-tasks together based on contextual cues and predicted future returns rather than relying on pre-defined skill boundaries.

Scalability to long sequences of interdependent goals (chained targets). The system maintains performance over horizons of 5+ sequential rearrangement objectives where error accumulation is high, demonstrating robustness against the degradation seen in task-specific imitation baselines.

Efficiency in data utilization by learning directly from sub-task demonstrations instead of requiring expensive, fully labeled complete-task trajectories or oracle plans for every new task variation. This makes the system more practical for real-world deployment where comprehensive skill labeling is infeasible.

) Comparison to Existing AI Paradigms: Bridging the Gap

The improved system fills a critical gap between two existing AI paradigms:

It surpasses Task-Specific Imitation (like Skill Transformer) by overcoming its deterministic limitations in multimodal action spaces, allowing it to select among multiple valid actions based on predicted long-term value.

It outperforms Privileged Explicit Skill Systems (Oracle Planners) by achieving comparable or superior performance without the massive engineering overhead of manually defining a task plan and explicitly decomposing the skill sequence for every new environment configuration.

In summary, the improved AI system is a data-driven, implicit coordinator that can solve long-horizon robotic rearrangement by learning to adaptively coordinate multiple learned behaviors based on real-time context and value estimation, making it more generalizable and robust than current explicit skill or task-specific imitation methods.

Long-horizon robotic rearrangement is commonly formulated as a skill-sequencing problem, where distinct behaviors are explicitly represented and coordinated by a planner or high-level policy. We investigate whether such explicit behavior identities and sequencing interfaces are necessary at all. We introduce implicit behavior coordination from sub-task demonstrations, where separately collected behaviors are coordinated without behavior identity labels, complete-task demonstrations, or task-ordering supervision. Our key observation is that overlap between sub-task demonstrations induces multimodal action distributions that need not be resolved through explicit behavior partitioning. Instead, this overlap-induced multimodality can be exploited as a coordination resource. We instantiate this idea with a shared Flow Matching policy that preserves multiple action modes and critic-guided in-sample planning that propagates task value across demonstrations and selects task-relevant modes. Experiments in Habitat and on a real robot show that implicit behavior coordination remains effective under reduced cross-behavior overlap, larger behavior mixtures, longer horizons, and execution failures, supporting the idea that long-horizon coordination can emerge directly from sub-task demonstrations without explicitly recovering or sequencing behavior identities.