$χ_{0}$: Resource-Aware Robust Manipulation via Taming Distributional Inconsistencies

Abstract: High-reliability long-horizon robotic manipulation has traditionally relied on large-scale data and compute to understand complex real-world dynamics. However, we identify that the primary bottleneck to real-world robustness is not resource scale alone, but the distributional shift among the human demonstration distribution, the inductive bias learned by the policy, and the test-time execution distribution -- a systematic inconsistency that causes compounding errors in multi-stage tasks. To mitigate these inconsistencies, we propose $χ{0}$, a resource-efficient framework with effective modules designated to achieve production-level robustness in robotic manipulation. Our approach builds off three technical pillars: (i) Model Arithmetic, a weight-space merging strategy that efficiently soaks up diverse distributions of different demonstrations, varying from object appearance to state variations; (ii) Stage Advantage, a stage-aware advantage estimator that provides stable, dense progress signals, overcoming the numerical instability of prior non-stage approaches; and (iii) Train-Deploy Alignment, which bridges the distribution gap via spatio-temporal augmentation, heuristic DAgger corrections, and temporal chunk-wise smoothing. $χ{0}$ enables two sets of dual-arm robots to collaboratively orchestrate long-horizon garment manipulation, spanning tasks from flattening, folding, to hanging different clothes. Our method exhibits high-reliability autonomy; we are able to run the system from arbitrary initial state for consecutive 24 hours non-stop. Experiments validate that $χ{0}$ surpasses the state-of-the-art $π{0.5}$ in success rate by nearly 250%, with only 20-hour data and 8 A100 GPUs. Code, data and models will be released to facilitate the community.

• The paper presents χ₀, a framework integrating Model Arithmetic, Stage Advantage, and Train-Deploy Alignment to bridge distributional gaps in robot learning.
• It demonstrates nearly 250% improvement over state-of-the-art baselines using moderate resources and targeted recovery strategies.
• The approach provides a formal methodology for aligning training, inference, and recovery phases in dual-arm garment manipulation.

$χ_{0}$: Resource-Aware Robust Manipulation via Taming Distributional Inconsistencies

Introduction

The paper "$χ_{0}$: Resource-Aware Robust Manipulation via Taming Distributional Inconsistencies" (2602.09021) presents a comprehensive framework that directly addresses the critical challenge of distributional inconsistencies across data collection, model training, and deployment in robotic manipulation. The authors identify that simply scaling data and compute is insufficient for achieving robust multi-stage, long-horizon manipulation in real-world settings due to three pivotal sources of distributional shift: (1) sparse training coverage, (2) inference-execution temporal mismatch, and (3) absence of recovery behaviors within training data. The $χ_{0}$ system leverages three synergistic strategies—Model Arithmetic (MA), Stage Advantage (SA), and Train-Deploy Alignment (TDA)—to systematically bridge these distributional gaps and enable high-reliability collaborative dual-arm garment manipulation within constrained resources.

Distributional Inconsistency: Formalization and Challenges

The paper formalizes robot learning as a sequence of distributional transitions:

• $P_\text{train}$: Distribution over expert human demonstrations.
• $Q_\text{model}$: Policy inductive bias learned from $P_\text{train}$.
• $P_\text{test}$: Deployed trajectory distribution, subject to inference-control latency and physical constraints.

Empirical deployment exposes three critical mismatches:

1. Coverage deficiency: $P_\text{train}$ under-represents the full task manifold, inducing narrow $Q_\text{model}$ bias.
2. Temporal mismatch: Latency and stage ambiguity produce misaligned execution, compounding errors across horizons.
3. Failure cascade: Training lacks recovery strategies; minor drifts at test time lead to unrecoverable failures.

   Figure 1: Pipeline of $χ_{0}$, capturing the progressive distributional alignment through MA, SA, and TDA across training, modeling, and deployment stages.

Methodology

Model Arithmetic (MA)

MA implements weight-space merging of independently trained policies on disjoint subsets of the demonstration data, thereby synthesizing a unified policy with expanded coverage. The interpolation coefficients are optimized using OOD validation loss, specifically on recovery trajectories collected via DAgger. The authors benchmark four merging heuristics—uniform averaging, inverse loss weighting, gradient descent, and greedy search—demonstrating that OOD validation yields greater robustness and statistical stability than in-domain splits.

Figure 2: Souping strategies in Model Arithmetic, showing diverse merging protocols with inverse loss weighting favoring low-loss checkpoints.

Stage Advantage (SA)

SA mitigates high-variance advantage estimation in long-horizon, multi-stage tasks. By decomposing tasks into semantic stages with explicit sub-goals, SA replaces noisy value differences with direct advantage prediction between state pairs, conditioned on the stage. This approach leverages a VLM-based model and generates stable binary progress signals for robust advantage-weighted behavior cloning.

Figure 3: Cumulative stage advantage signals for tasks A/B/C, illustrating recovery and failure structure with stage-conditioned value.

Train-Deploy Alignment (TDA)

TDA closes the loop between training and deployment via spatio-temporal data augmentation, heuristic DAgger, and a novel temporal chunk-wise action smoothing algorithm. Heuristic DAgger proactively collects recovery data from hand-designed failure states. The smoothing algorithm maintains temporal continuity in action execution, efficiently bridging inference-control latency without architectural modification.

Figure 4: TDA strategies with T-SNE visualizations showing increased distributional alignment at each stage.

Experimental Evaluation

Robot Setup and Task Suite

Collaborative dual-arm robots execute high-variance long-horizon tasks: T-shirt flattening/folding (Easy), conditional retrieval/sorting (Medium), and garment hanging (Hard). The evaluation protocol utilizes success rate, throughput, retry cost, and composite task score over 10 trials for three garment types, with average values and cumulative standard errors reported.

Figure 5: Experimental robot setup with dual-arm collaborative platforms.

System Efficacy

Experimental ablations verify that each module individually and in composite form provides monotonically increasing performance across metrics, with MA and SA delivering marked improvements in success rate and throughput, and TDA augmenting recovery but incurring higher retry cost.

Figure 6: Efficacy breakdown of the $χ_{0}$ system on Task A, highlighting synergistic effect of module integration.

Model Arithmetic Performance

MA consistently outperforms both single-best and full-data baselines across all tasks and metrics, with OOD validation demonstrating reduced standard error and increased reliability. Greedy search merging with DAgger validation loss emerges as the most effective.

Figure 7: MA ablations for Task C, showing statistically significant gains over baselines with greater throughput and stability.

Stage Advantage Ablation

SA demonstrates superior numerical stability to value-difference-based advantage estimation. Stage conditioning prevents spurious inactivity and minimizes dense retry costs in long-horizon, conditional tasks.

Figure 8: SA ablations for Task A/B, quantifying improved numerical stability and performance correlation.

Train-Deploy Alignment Ablation

Heuristic DAgger achieves substantial gains in failure recovery and overall performance, rivaling standard DAgger with greatly reduced operational overhead. Temporal chunk-wise smoothing consistently boosts inference performance compared to temporal ensembling and RTC across action parameterizations and task complexity.

Figure 9: DAgger ablations on Task A, showing the positive effect of both heuristic and standard interventions on throughput and recovery.

Figure 10: DAgger ablations on Task C, reinforcing trends seen on easier tasks.

Figure 11: Control strategy and augmentation ablations, with chunk-wise smoothing outperforming alternatives.

Module-Specific Ablations Across Action Representations

Evaluation of absolute and delta joint control schemes shows chunk-wise smoothing maintains its superiority in inference stability and task success, regardless of representation.

Figure 12: Task A control/augmentation ablations across action parameterizations.

Figure 13: Task C control/augmentation ablations, indicating consistent chunk-wise smoothing gains even in fine-grained tasks.

Figure 14: Task B control/augmentation ablations, demonstrating task-dependent performance but overall benefit of smoothing and RTC combinations.

Implications and Future Directions

The $χ_{0}$ framework sets an operational paradigm for resource-efficient robust manipulation through targeted distributional alignment at all phases, explicitly overcoming compounding failure modes endemic to physical robot deployment. The empirical results are highly significant: nearly 250% improvement over state-of-the-art baselines with less than one day of demonstration data and moderate computational resources. The findings underscore the critical role of OOD validation in policy merging, stage-conditioned advantage in stable supervision, and proactive failure recovery in sustaining autonomous operation.

Practical implications extend to industrial long-horizon manipulation, where throughput and reliability are paramount and data collection is costly. The methods are directly extensible to non-garment, contact-rich manipulation domains and could facilitate generalist robotic foundation models when combined with scalable MA across heterogeneous tasks.

Theoretical implications include a formalization of robot learning as alignment of three distinct distributions, motivating future research in efficient data utility metrics, integration of pre-trained priors, and unsupervised advantage modeling. Limitations include scalability to rigid-body tasks and absence of persistent prior retention analysis.

Conclusion

$χ_{0}$ represents a rigorously engineered, modular system for resource-aware robust manipulation. By systematically taming coverage, temporal, and recovery inconsistencies with MA, SA, and TDA, it achieves unprecedented reliability in collaborative long-horizon garment manipulation within strict resource constraints. The results establish a methodological foundation for future robotic policy alignment and sustained autonomous operation in unstructured real-world environments.