Thank you very much for your constructive feedback and acknowledgment of our efforts. Following are our responses to all your concerns.

Skill Extraction need addtional training cost to mitigate the skill imbalance problems.

As detailed in Appendix B.2, we ensure that the total number of training iterations remains consistent across all methods. Specifically, the sum of stage 1 iterations (Skill Extraction) and stage 2 iterations (Skill Enhancement and Skill-Based Policy Learning) in GO-Skill is equal to the total iterations of the baselines. Importantly, since Skill Enhancement and Skill-Based Policy Learning are entirely independent learning processes, they are trained in parallel during the second stage of GO-Skill. As a result, no extra training time or iterations are introduced compared to the baselines. This design allows us to incorporate skill balancing mechanisms without increasing overall computational cost or compromising fairness in comparison.

Over-reliance on MetaWorld limits domain diversity, such as LIBERO.

Thank you for suggesting the LIBERO benchmark. However, we would like to clarify that LIBERO is fundamentally a multi-task imitation learning benchmark rather than a MTRL benchmark. The dataset only provides a limited number of expert demonstrations (50 trajectories per task) and lacks reward information (only a binary success signal). This setup does not fully capture one of the key advantages of GO-Skill: suboptimal trajectories from certain tasks can still facilitate skill learning and potentially benefit other tasks leveraging the same skill. We have analyzed this phenomenon in Section 5.4 of the main paper.

Nevertheless, we adapted our approach to the LIBERO setting by employing an imitation learning paradigm, removing the return-conditioning in DT-based models, and comparing GO-Skill, PromptDT, and MTDT under this framework. Additionally, since LIBERO is a vision-based environment, we utilized R3m[1] to encode image observations into vector representations. We use the LIBERO90 setting, and the experimental results are as follows:

From the experimental results, it can be seen that the explicit knowledge sharing provided by skill abstraction can also help agent to learn multiple tasks quickly in a multi-task imitation learning scenario.

Furthermore, based on suggestions from other reviewers, we also conducted few-shot experiments on Cheetah-Vel and Ant-Dir, following the PromptDT setup. The results are shown below:

As shown in the table, GO-Skill achieves comparable performance to PromptDT. This is expected because these environments are relatively simple, and the high similarity between different tasks causes both GO-Skill and PromptDT to reach the performance ceiling of these benchmarks.

[1] Nair, Suraj, et al. R3m: a universal visual representation for robot manipulation. CoRL 2022.

What does "dynamic transfer" mean in Section 3.1.

"Dynamic transfer" refers to the MDP’s transition dynamics, i.e., how the state evolves. The key distinction here is that this dynamic transition describes the state change after H steps of decision-making, rather than a single-step transition. If this phrasing has caused any misunderstanding, we will revise it to "H-step dynamic transfer" in the revised manuscript to clarify the meaning.

Thank you once again for your valuable feedback. If you need further elaboration or additional points to include in the response, we welcome further discussion to ensure everything is clear and satisfactory.

Thank you very much for your constructive feedback and acknowledgment of our efforts. Following are our responses to all your concerns.

Perform on locomotion tasks "Cheetah-vel" and "Ant-dir"

We followed the experimental setup of PromptDT and conducted few-shot experiments on the Cheetah-Vel and Ant-Dir environments. The detailed results are as follows:

As shown in the table, GO-Skill achieves comparable performance to PromptDT. This is expected because these environments are relatively simple, and the high similarity between different tasks causes both GO-Skill and PromptDT to reach the performance ceiling of these benchmarks.

Furthermore, based on suggestions from other reviewers, we also conducted experiments on another multi-task imitation learning benchmark LIBERO. Due to space constraints, we kindly refer you to our response to Reviewer FkAC for detailed information. For convenience, we recommend using Ctrl+F to search for "LIBERO benchmark" to quickly locate the relevant response.

The trajectory-level state difference representation of skills may not capture all relevant aspects of useful skills.

We appreciate your thoughtful observation. Indeed, no single representation can capture all aspects of useful skills. GO-Skill adopts a goal-based representation to define skills in a task-agnostic manner. As shown in our ablation study in Section 5.3, this representation outperforms action-based alternatives. We believe this is because multiple action sequences can lead to the same state transition, and abstracting skills based on goals allows the agent to generalize across these variations, leading to more transferable and reusable skills.

We agree that goal-based representation may not always be optimal, but our framework is flexible and can accommodate alternative skill representations. We thank you for highlighting this direction, and we plan to further explore more expressive representations as part of our future work.

How GO-Skill could be extended to learn skills with different temporal horizons?

Thank you for raising this important point. While the current GO-Skill implementation adopts a fixed temporal horizon for simplicity and consistency, the framework itself is fully compatible with variable-length skills.

To support variable-length skills, the primary change would be in the skill decoder, which can be extended to learn a terminate signal to determine when to exit a skill. The key challenge lies in how to extract and define variable-length skills during the offline skill discovery phase. As mentioned in our Future Work Section, this is a central focus of our ongoing research. We are currently exploring approach by applying thresholds on state changes to identify skill. When the state change exceeds a certain threshold, it indicates the need to switch to a new skill.

We believe enabling dynamic skill lengths will further enhance the expressiveness and adaptability of GO-Skill, and we look forward to sharing these results in future work.

Add a reference to Appendix C in results section.

Thank you for your suggestion and for recognizing the value of the visualization results. We will add a reference to Appendix C in the results section in the revised manuscript.

"get prompt" appears outside the "for each task" loop in Alg.2.

Thank you for pointing out this mistake. This is indeed a writing error, and we will correct it in the revised manuscript.

Additionally, to clarify the prompt selection process: similar to PromptDT, we construct a Prompt Set for each task with the top-N highest return trajectories (N=4 in our experiments). Each time "get prompt" is called, a trajectory is randomly sampled from the corresponding task’s Prompt Set.

Goal encoder should be $\mathcal{G}: \mathcal{S} \rightarrow \mathcal{Z}$

Thank you very much for pointing out the error. We will correct it in the revised manuscript.

Thank you once again for your valuable feedback. If you need further elaboration or additional points to include in the response, we welcome further discussion to ensure everything is clear and satisfactory.

Thank you very much for your constructive feedback and acknowledgment of our efforts. Following are our responses to all your concerns.

Overlook offline MTRL baseline CDS.

Thank you for pointing out the omission of the CDS, which is indeed a significant contribution to offline MTRL. We will cite and discuss CDS more thoroughly in the revised manuscript.

CDS focuses on a conservative data-sharing strategy, selectively utilizing data from other tasks to benefit learning on a given target task. In contrast, GO-Skill approaches knowledge sharing at a higher level of abstraction by discovering reusable skills that can be naturally composed across tasks. Moreover, these two methods could be integrated, where the high-level skill-based policy in GO-Skill can be enhanced by applying conservative filtering to determine which skill-based transitions to share across tasks as proposed in CDS, leading to more robust policy learning. We believe this hybrid direction holds promise and represents an exciting avenue for future work.

Unfortunately, CDS’s codebase and experimental data are not publicly available, making a direct comparison difficult. Moreover, reimplementing such a method without official reference results for validation could risk an unfair or inaccurate evaluation. Nonetheless, we are currently emailing the authors to request the necessary resources and guidance for reproduction. If successful, we will experimentally explore the potential complementarity between CDS and GO-Skill, and we are excited to investigate the results in the final version.

Only a single environment setup is used — Metaword.

Based on suggestions from other reviewers, we have conducted experiments on two additional benchmarks beyond MetaWorld. One is the locomotion task suite (Cheetah-Vel and Ant-Dir) used in PromptDT, and the other is LIBERO, a multi-task imitation learning benchmark.

Due to space constraints, we kindly refer you to our response to Reviewer FkAC for detailed information. For convenience, we recommend using Ctrl+F to search for "LIBERO benchmark" to quickly locate the relevant response.

The method is quite complex and it is not clear whether all the components are necessary in general. I'd suspect for different MTRL task suites, not all the proposed contributions would be necessary.

As noted above, we also evaluate GO-Skill on additional benchmarks. The results demonstrate the general applicability of GO-Skill, supporting that it is not tailored to a specific environment.

In addition, our method is not conceptually complex. It consists of two main components: the skill model, which is shared across all tasks and generates actions to interact with the environment, and the skill-based policy, which completes specific tasks by composing different skills. In our ablation study, we systematically analyze both the skill model architecture and the methods for handling skill imbalance, demonstrating that each component contributes to overall performance.

Lacks success rate, which limits the comparability of the results.

We chose to report return as the primary evaluation measure in MetaWorld because it provides a more informative metric of task performance. For example, in the 'drawer-close' task, the environment considers the task successful once the drawer is closed past a certain threshold — even if it is not fully closed. However, the agent can continue to receive additional rewards for closing the drawer further, indicating a higher degree of task completion. In such cases, success rate acts as a binary threshold, while return reflects the quality and completeness of the agent’s behavior.

In addition, we conducted additional experiments and report the success rate of GO-Skill as follows:

How is the hierarchical skill abstraction fundamentally different from previously proposed methods?

Our skill-based MTRL framework offers a fundamentally different perspective from prior methods. Specifically, most prior multi-task RL methods achieve generalization through implicit knowledge sharing through various parameter-sharing strategies. In contrast, GO-Skill adopts an explicit knowledge sharing approach by directly extracting reusable skills from offline data. A key advantage of this explicit skill abstraction is its ability to better leverage suboptimal data. In offline settings, a suboptimal skill in one task may serve as an optimal skill for another, allowing GO-Skill to effectively master useful behaviors across tasks. This leads to more efficient skill acquisition and enhances generalization across diverse multi-task scenarios.

Thank you once again for your valuable feedback. If you need further elaboration or additional points to include in the response, we welcome further discussion to ensure everything is clear and satisfactory.

Thank you very much for your constructive feedback and acknowledgment of our efforts. Following are our responses to all your concerns.

Technical components of the paper are quite standard, so better justification would be beneficial.

We appreciate your comment and acknowledge that certain technical components of our framework, such as vector quantization and transformer-based policy modeling, are standard techniques. However, these choices are made to validate the effectiveness of our framework, rather than serving as core contributions. Alternative implementations, such as diffusion-based policies or different discretization methods, could also be applied without altering the fundamental ideas of our approach.

The key contribution of GO-Skill lies in proposing a novel skill-based MTRL framework that enables explicit knowledge sharing through skill learning and composition, inspired by how humans generalize through modular skills. Furthermore, we introduce Goal-Oriented Skill Abstraction, a new method for defining and extracting reusable skills based on goal transitions rather than action sequences or rewards. This task-agnostic representation enhances skill transferability and enables effective skill discovery even from sub-optimal data.

In Section 5, we justify different components through an ablation study. The results show that while standard techniques like VQ contribute to performance improvements, Goal-Oriented Skill provides the most significant boost, highlighting its critical role. In addition, all ablated variants of GO-Skill outperform the baselines, demonstrating the effectiveness of our skill-based framework.

We hope this clarifies that our contributions are primarily conceptual and architectural, and that our use of standard components supports, rather than limits, the novelty of our approach.

The main comparison results shown in Fig. 5 are presented without actual numbers. The y-axis starting from 1.5 makes it difficult to judge the difference of different settings.

We appreciate your valuable feedback on Fig.5. In the revised manuscript, we will include a table presenting the exact numerical results corresponding to Fig.5 for clarity. Below is the table summarizing the results:

The reason for setting the y-axis starting from 1.5k is to better highlight the performance differences among different methods while maintaining a consistent y-scale across all four settings for easy cross-comparison. In addition, we would greatly appreciate any suggestions you may have for an improved visualization standard, and we will gladly incorporate the necessary adjustments in the revised manuscript accordingly.

Thank you once again for your valuable feedback. If you need further elaboration or additional points to include in the response, we welcome further discussion to ensure everything is clear and satisfactory.