COOPERA: Continual Open-Ended Human-Robot Assistance

Thank you for your insightful and constructive comments. We appreciate your recognition of the contribution of our framework and the introduced benchmark, as well as the design of our ablation study. This recognition is also kindly shared by other reviewers (Reviewer kYPv, Reviewer Zf7z, Reviewer irvc). In response to your suggestions, we conducted additional experiment to strengthen our analysis. For all HRI-related experiments, we report results from our main method under the more challenging collaboration type 2 (one trial per setting) due to time constraints. Below, we address each of your concerns.

Q1: The discussion on the benchmark problem is not sufficient, given it is a newly introduced testbed.

A1: Thank you for the valuable suggestions! We fully agree that providing additional explanations to Section 3.1 is essential, given that COOPERA is a newly introduced testbed. Central to COOPERA is our proposed human model and the continuous feedback mechanism operating at a daily scale. With these components, COOPERA differs from previous HRC benchmarks [1, 2] in two aspects:

1. Existing benchmarks typically focus on short-term, episodic HRC tasks that are closed-set and predefined. In contrast, our human model generates open-ended, environment-driven, and dynamic human activities that vary with time, enabling COOPERA to support long-horizon collaboration tasks spanning multiple days.

2. The daily feedback mechanism allows robots to continuously optimize their behavior over extended periods, leading to increasingly personalized interactions tailored to individual human traits and preferences.

We will explicitly add these points to Section 3.1.

Q2: The comparison with baseline methods needs more discussions.

A2: Thank you for highlighting this! We strongly agree that additional detailed analysis on the performance of our method vs. the baselines are needed. For paper writing, we will include more analysis in Section 4.3 and move some of the implementation details in Section 4.1 to Appendix.

Below we provide an in-depth discussion of why each method achieves distinct results. We evaluate COOPERA against six baselines: (1) Direct Prompting, (2) Direct Finetuning, (3) GT_Intention, (4) Random, (5) Intention Agnostic, and (6) Human & Context Agnostic.

For within-day improvement (Fig. 6a), our method achieves the highest performance gain by explicitly modeling the correlation between human intentions/tasks, traits, and temporal dependencies. As the day progresses, the robot accumulates interaction history and adapts its behavior to the human preferences. (1) Direct Prompting shows minimal improvement as it depends on retrieved interaction history without learning the human preferences. (2) Direct Finetuning performs slightly better but still struggles due to its rigid mapping from inputs to intentions/tasks, making it biased toward frequent training examples and overlooking the varying distribution of human behavior. (3) GT_Intention, despite having ground-truth intentions, shows limited gains because receiving the correct intention upfront removes the need to model temporal context or adapt over time. Its performance is static by design and it serves as the upper bound method. (4) Random degrades over time due to the absence of validation. (5) Intention Agnostic shows moderate improvement but lags behind our method, as it skips intention inference and thus loses contextual information, which is problematic since temporal dependencies between intentions are typically stronger than between tasks (validated in Section 4.2). (6) Human & Context Agnostic captures only superficial time-task correlations and ignores human traits or temporal context, leading to weak within-day gains.

Across-day improvement (Fig. 6b) further highlights the importance of learning the correlation between human intentions, human profile, and temporal dependencies. Our method achieves strong across-day gains—ranking second only to (3) GT_Intention, which receives the correct daily intention and thus bypasses the most challenging aspect of the problem. (1) Direct Prompting, (2) Direct Finetuning, and (4) Random exhibit minimal across-day improvement for the same reasons they underperform within a day: lack of learning from feedback, rigid input-behavior mappings, or absence of validation. (5) Intention Agnostic and (6) Human & Context Agnostic show some improvement, but their performance remains significantly lower than our main approach, as they exclude critical components such as explicit intention modeling and the integration of human profile and temporal context.

Q3: What is the baseline method used in Table 2?

A3: Thank you for raising this point! We will add more explanations regarding the baselines for this experiment for better clarity. Table 2 evaluates our framework’s out-of-domain generalization (Section 4.3, line 319) in two scenarios: (1) scene generalization and (2) human generalization.

In the scene generalization scenario, we first train the robot to collaborate with the same human across four distinct scenes (as defined in Setting 2: same human, different scenes). We then evaluate its ability to generalize to a fifth, unseen scene with the same human. The baseline here is the robot’s average performance during the initial interaction in the unseen scene without any finetuning on the previous four scenes, evaluated across 10 different humans.

In the human generalization scenario, we train the robot to collaborate with three different humans in the same environment (Setting 3: different humans, same scene). We then assess its generalization to a fourth, previously unseen human. The baseline in this case is the robot’s initial, unadapted performance when interacting with the new human, without finetuning on the previous three, averaged across 5 new humans.

Notably, our method outperforms the baseline in both cases, indicating that the robot learns transferable representations of human traits and environments. This shows that the system avoids overfitting and generalizes well to new humans and scenes.

Q4: How do you define “success” for each task?

A4: Thank you for pointing this out! We agree that evaluation clarity is important and will revise Section 4.3 (line 289) accordingly. In COOPERA, we assess F1-based success rates using three complementary methods, consistent with existing HRC benchmarks [1, 2, 3].

In our framework, the robot-VLM predicts a superset of tasks to assist the human. An LLM-based classifier then filters this list, assigning "yes" or "no" to each task. For baselines that lack this component (e.g., Direct Prompting or Fine-tuning), we assume all predicted tasks are labeled "yes". Please see Section 3.3 for details.

We measure success using three criteria:

1. Predicate-based success: Tasks are executed in the simulator and evaluated via predefined predicate functions (Appendix D.3, line 714). Matching is based on object classes, following Watch-and-Help [1], and F1 is computed over these outcomes.

2. LLM-based success: The simulated human (human-LLM) is shown the robot’s predicted tasks and the ground-truth intention, and gives binary ("yes/no") feedback on whether each task fulfills the intention (please see Appendix C.3 and F for detailed feedback machnism and exact prompt). F1 is computed from this feedback.

3. Human verification: Real humans replace the simulated human in (2) and provide judgments. This ensures our evaluation extends from simulation to real-world validation by human evaluators.

Q5: Is there any visualization or video demonstration showing how the trained robot agent performs during testing?

A5: Yes! For the predicate-based evaluation, the robot directly executes tasks in the simulated environment. We also provide functions for the human agent to perform actions collaboratively when the robot’s assistance is approved. This enables us to create visualizations and demonstration videos illustrating the human-robot collaboration process, as in Fig. 7 of our paper. Due to this year's conference policy restricting attachment of supplementary PDF during rebuttal, we are unable to include these visuals here. However, we will add additional visualization examples to the Appendix and our project website after the rebuttal period. Furthermore, we will release all code related to our framework.

Q6: While the proposed method gains the best improvement over time, it seems that it does not converge to the one with ground truth intentions, as in Fig. 6.

A6: This is a great question! We conduct one additional experiment. We extend the current setup by doubling the collaboration duration to observe continued adaptation trends.

From the results, we observe that performance stabilizes around day 8. While our benchmark focuses on collaboration up to day 5, our main interest lies in the personalization trend. Simulating longer HRC is an easy extension within our framework. We believe this stabilization reflects a natural saturation point. By day 8, most human preferences driven by traits and routines are learned, allowing the robot to form a reliable alignment. Given the capacity of our 7B model and LoRA-based finetuning, further gains become marginal without introducing more capable models or finer-grained learning signals such as multimodal cues or reasoning chains.

References:

[1] Watch-and-help: A challenge for social perception and human-ai collaboration. ICLR 2021.

[2] Partnr: A benchmark for planning and reasoning in embodied multi-agent tasks. ICLR 2025.

[3] NOPA: Neurally-guided Online Probabilistic Assistance for Building Socially Intelligent Home Assistants. ICRA 2023.

Thank you for your insightful and constructive comments. We appreciate your recognition of the originality and the contribution of our framework, the novelty of our human simulation pipeline, and the comprehensiveness of our experimental design. This recognition is also kindly shared by other reviewers (Reviewer kYPv, Reviewer Zf7z, Reviewer PjNb). In response to your suggestions, we conduct additional experiments to strengthen our analysis. For all HRI-related experiments, we report results from our main method under the more challenging collaboration type 2 (one trial per setting) due to time constraints. Below, we address each of your concerns.

Q1: How sensitive is the COOPERA pipeline to the choice of LLM/VLM backbone and how would models of different sizes impact the performance? What is the reason behind the current choice?

A1: We choose LLaMA-3.2-11B as the robot-VLM and Mistral-7B-Instruct-v0.2 as the robot-LLM classifier for two reasons. First, both are open-source under research-friendly licenses, offering strong performance, transparency, and reproducibility for benchmarking. In contrast, proprietary models like GPT require costly API access, are frequently updated, and make results harder to reproduce. Second, 7B–11B models strike a practical balance between performance and computational efficiency [1, 2]. They support single-GPU inference (24GB RAM) and LoRA finetuning with just 2–3 GPUs, making them accessible to academic labs and the broader research community.

Yet, we fully agree it is important to evaluate the framework’s robustness to different VLM backbones. We conduct two additional experiments by replacing the robot-VLM with a stronger model (GPT-4o) and a weaker one (Qwen2.5-VL-3B-Instruct).

As expected, stronger models produce better performance, aligning with findings in Embodied Agent Interface [3], a benchmark that tests the impact of LLM capacity on LLM-powered embodied AI systems. Notably, the performance gap between GPT-4o and LLaMA is larger than that between LLaMA and Qwen. However, smaller VLMs like Qwen still achieve reasonable performance at the end of day 5 and show a clear personalization trend. This demonstrates the robustness of our framework across varying model capacities.

Q2: The method uses specific LLM prompts and parameters. Deeper analysis is needed.

A2: Thank you for pointing this out! We agree that prompt design and parameter tuning are important considerations for any LLM-based system. To maintain full control and interpretability of our framework, we handwrote our prompts. However, we also acknowledge there are open-source tools that automate prompt optimization. To ensure our performance is not the result of over-engineered prompts, we conduct an additional experiment by integrating DSPy [4, 5], an open-source framework for self-composing and self-improving prompts, into our assistive robot agent. DSPy allows prompt tuning using only high-level instructions and input-output examples. We evaluate our system with DSPy-generated prompts.

We observe a slight performance gain using DSPy. DSPy leverages the OPRO algorithm [8], and its evaluation reports an average gain of 5% in expert domains. In our case, we see a 3% improvement, which we find reasonable. It suggests that our hand-written prompts are already well-designed, though not overly engineered.

Q3: Can COOPERA be extended to support real-time, in-episode adaptation (e.g., continual feedback during a task)? What components of the current system would need to be modified to support this?

A3: This is a great question! COOPERA currently focuses on day-level adaptation, which we believe better reflects real-world usage—robot is optimized overnight while humans rest. However, enabling real-time adaptation through hour-level human/between-task feedback and developing corresponding online learning algorithms is a promising direction and a fundamental challenge for any LLM-based agent.

To conduct a preliminary study for this extension, we conduct one additional experiment.

COOPERA with hour-based communication: The human’s hourly intention/task is influenced by the robot’s action in the previous hour. For example, if the robot fails to clean the living room, the human may skip yoga to finish cleaning. We modify the human prompt to include robot's success/failure. We also ask the human to provide hourly feedback with reasoning to the robot (e.g., I disapprove this task because...). This helps the robot understand its mistakes and adjust its next action accordingly in the next hour.

The results are interesting. We expect higher or comparable performance due to frequent feedback but observe a slight drop. We find that the robot often repeats the previous intention to “make amends” when the human disapproves, while the simulated human sticks to its traits and only notes the error in feedback. Extending COOPERA to support more frequent feedback and developing new assistive strategies are promising directions for future work.

Q4: Given the long-term nature of the simulation and human profiles, are there qualitative or longitudinal metrics that could better capture: 1) The robot’s improvement over time? 2) The perceived usefulness or support quality from the human’s perspective?

A4: Thank you for the suggestion! Our current evaluation focuses on quantitative performance (predicate-based, LLM-based, and human-verified success rates), following conventions in existing HRC benchmarks [6, 7]. We fully agree that qualitative and longitudinal metrics can better capture perceived usefulness and long-term improvement.

We propose two directions for future work:

1. Subset-Based Evaluation for Intention Alignment Over Time: At each hour, the robot infers the human intention and proposes a task set. Our current LLM-based evaluation only gives binary approval/disapproval to each task in the set. We can extend this by selecting subsets of the robot’s proposed tasks and asking the simulated human whether each subset fulfills their intention. Ideally, some subsets will be approved, others rejected. By recursively forming and evaluating the subsets, we can determine which combinations of tasks are most critical for intention fulfillment. We can then assign a final score to the original full set by summing the scores of approved subsets (+1 each) and disapproved subsets (-1 each).

2. Human Preference Between Successive Robot Policies: For HRC at day t = n, we can compare a robot finetuned on data from t = 0 to n-1 versus t = 0 to n-2, then ask the simulated human to express a preference. This offers insight into how alignment improves across days.

References:

[1] Grattafiori, Aaron, et al. The Llama 3 Herd of Models. ArXiv, 2024.

[2] Kaplan, Jared, et al. Scaling Laws for Neural Language Models. ArXiv, 2020.

[3] Li, Manling, et al. Embodied Agent Interface: Benchmarking LLMs for Embodied Decision Making. NeurIPS, 2024.

[4] Khattab, Omar, et al. DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines. ArXiv, 2023.

[5] Xian, Jasper, et al. Prompts as Auto-Optimized Training Hyperparameters: Training Best-in-Class IR Models from Scratch with 10 Gold Labels. ArXiv, 2024.

[6] Puig, Xavier, et al. Watch-and-help: A challenge for social perception and human-ai collaboration. ICLR, 2021.

[7] Chang, Matthew, et al. Partnr: A benchmark for planning and reasoning in embodied multi-agent tasks. ICLR, 2025.

[8] Yang, Chengrun, et al. Large language models as optimizers. ICLR, 2024.

We are happy our replies addressed your questions! Thank you for your kind response and for improving the score from 4 (Weak Accept) to 5 (Accept) in support of our work. We will add the additional experiments to the paper or Appendix with clear references.

Thank you for your insightful and constructive comments. We appreciate your recognition of our framework’s contribution, careful integration of foundation models, and completeness of our evaluation. This recognition is also kindly shared by other reviewers (kYPv, irvc, PjNb). Following your suggestions, we added new experiments to strengthen our analysis. For HRI experiments, we report results under the more challenging collaboration type 2 (one trial per setting) due to time constraints. Below, we address each of your concerns.

Q1: The evaluation relies heavily on synthetic human simulations, with limited validation against real human behavior.

A1: Thank you for pointing this out! COOPERA relies on synthetic human simulations due to ethical, safety, and practical constraints. Our goal is to study long-term, open-ended HRC, which better mirrors real-world use, unlike most HRC benchmarks focused on episodic, closed-set tasks [1–3]. To support long-horizon HRC, we need humans who can continuously collaborate with robots and provide feedback. However, running full real-world studies with physical robots in household over multiple days is costly and raises ethical and safety concerns. As an initial step, we build a scalable simulation pipeline enabling personalized, long-term HRC entirely in simulation (supported by Reviewer kYPv, irvc).

To assess the realism of our simulated human model, we present extensive analysis in Section 4.2 (also noted by Reviewer irvc), including classifier-based evaluations (distinguishability, diversity, psychometric coherence, temporal dependence) and two user studies. However, we fully agree more validation is necessary.

To address your concern, we conduct a new experiment: six users provide traits and daily intentions over 5 days. Using the same traits, we prompt the LLM to simulate intentions for the same period. Both sets are aggregated into single paragraphs (removing time formatting like “9am:” to avoid inflated similarity), and semantic similarity is computed using SBERT (all-mpnet-base-v2) and OpenAI embeddings (text-embedding-3-small). We compare against three baselines: (1) prompting without human profile (generic), (2) mismatched LLM-human intention pairs, and (3) generating all intentions at once without hour-level grounding (as in Table 4 ablation).

From the results, Generic and Misaligned yield moderate similarity ($\sim$0.5), since sentence encoders naturally assign partial similarity to structurally similar content. Our main method achieves notably higher scores, indicating closer alignment with human-specific intent. Sentence-BERT yields slightly higher similarity than OpenAI embeddings, likely due to its objective being explicitly tuned for sentence-level semantic comparisons.

Q2: Notable dependence on foundation models throughout the pipeline.

A2: Great point! Both our human simulation pipeline and benchmark methods rely on foundation models. As COOPERA extends closed-set tasks [1, 3] to open-ended ones, foundation models are a natural choice for agent reasoning and decision-making [3], a trend also seen in open-ended embodied AI systems [4].

Yet, we fully agree on concerns with heavy reliance on LLMs, especially regarding pipeline brittleness/robustness and prompt over-engineering.

To address the concerns, we run three new experiments: 1) We replace the robot-VLM with a stronger model (GPT-4o) and 2) a weaker one (Qwen2.5-VL-3B-Instruct). 3) Test prompt sensitivity by swapping our handwritten prompts with DSPy [5], an open-source framework for prompt automation using only high-level instructions and I/O examples.

As expected, stronger models yield better performance [3]. Notably, smaller VLMs like Qwen still achieve reasonable performance by day 5 and show a clear personalization trend, demonstrating the robustness of our framework across model scales. We also observe a 3% gain with DSPy, compared to its reported result (5%) in expert domains. This suggests our handwritten prompts are well-structured but not over-engineered.

Q3: The use of synthetic data without substantial real-world validation raises concerns about completeness and the risk of “echo chamber” effects. Real-world applicability remains unclear.

A3: Thank you for bringing this up! The ultimate goal of COOPERA is to develop robot agents that assist real humans by adapting to their preferences over time. To validate real-world applicability (Section 4.4), we conducted human verification and collaborated with offline real humans. While our experiments are in simulation, the robot’s embodiment and actions match the real Fetch robot’s default settings in Habitat [7], enabling seamless transfer of generated action sequences to the real robot.

However, we fully agree more validation of real-world applicability is necessary. To address this, we include a new experiment. Three real humans replace the LLM and collaborate under setting 2. Users propose intentions based on their traits and adjust to robot predictions in real time. Users are shown retrieved object/motion sets and choose which to interact with and use. For simplicity and due to time constraints, users input responses as text instead of using keyboard to control the simulated agent.

The results are interesting. We expect lower performance due to real human variability, yet results are higher. After discussing with participants, we find they tend to recall and follow their mid-week routines, which are relatively stable. This consistency leads to lower variability than our simulated humans, allowing the robot to personalize more quickly and validating the behavioral diversity in our simulated humans.

Q4 & 5: Clarification on how success is defined and how feedback is provided to the robot.

A4 & 5: Thank you for pointing this out! Evaluation clarity is crucial and we will revise Section 4.3. We compute F1-based success using three methods, consistent with prior HRC benchmarks [1–3]. The robot’s LLM classifier outputs a yes/no-labeled task list, and F1 is computed using the following GT labels:

1. Predicate-based: Tasks are executed and evaluated by predicate functions (Appendix D.3), with class-based object matching following Watch-and-Help [1].

2. LLM-based: Simulated human judges robot tasks and gives binary feedback on if each task fullfills its intention (Appendix C.3 and F).

3. Human verification: Same as (2) but real humans provide the binary feedback.

For more details, please see our reply to Reviewer PjNb Q4 due to word limit.

Q6: In Fig. 6b, intent-agnostic and human-agnostic baselines show comparable learning progress to full proposed approach.

A6: Great question. Among six baselines (Section 4.3), Intention Agnostic is closest to our full method. It removes intention inference so tasks are predicted without modeling intention. Human & Context Agnostic learns only time-to-task mappings, ignoring traits and temporal cues.

In Fig. 6b, Intention Agnostic and Human & Context Agnostic reach $\sim$0.4 and $\sim$0.35 F1, while our full method achieves $\sim$0.5. Though the gap seems small due to averaging across all settings, Appendix Fig. 12 reveals: (1) Our method significantly outperforms both in harder settings (3 & 4, different humans), highlighting the value of intention modeling and personalization. (2) In easier settings (1 & 2, same human), both baselines perform similarly since trait learning matters less. (3) In multi-human settings (3 & 4), Intention Agnostic surpasses Human & Context Agnostic, showing while intention modeling helps, learning human-specific behavior is crucial for generalization.

Q7: How are robot actions defined, and what input modalities support complex task inference?

A7: Robot actions are implemented as primitives from Habitat’s task_action registry, combining navigation EE control: (1) MoveEEAction moves the EE via Cartesian steps with IK; (2) PickObjIdAction grasps objects by snap within a threshold; (3) PlaceObjIdAction releases via desnap; (4) ResetEEAction resets the EE to default. Navigation uses classical planning with AABB thresholds. These form a full pick-and-place pipeline. We will release all code.

For intention discovery, the robot-VLM receives five 1024×768 RGB frames of the human’s first task in collab type 2. In type 1, we also provide texts (Section 3.1). Intention-to-task decomposition is text-based. For complex intentions (e.g., “meditate”), we let VLM reason over visual cues, prior actions, and traits. For complex tasks (e.g., “brainstorm”), we use the VLM’s proposed task superset, filtered by our LLM-based classifier, to capture valid actions.

Q8: Paper writing. The structure limits the nuanced discussion of results and findings. Fig. 1 not referenced.

A8: Thanks for pointing out! For revision, we will: (1) move selected implementations from Sections 3.2 and 4.2 to Appendix; (2) expand discussion of findings, including new results addressing Q1–Q3 and Q6; and (3) clarify how success rate and feedback are defined, either in the main text or via Appendix references. We will also explicitly describe Fig. 1 in intro.

[1] Watch-and-help. https://arxiv.org/abs/2010.09890

[2] Partnr. https://arxiv.org/abs/2411.00081

[3] NOPA. https://arxiv.org/abs/2301.05223

[4] Embodied Agent Interface. https://arxiv.org/abs/2410.07166

[5] DSPy. https://arxiv.org/abs/2310.03714

[6] Habitat 2.0. https://arxiv.org/abs/2106.14405

Thank you for your insightful and constructive comments. We appreciate your recognition of the novelty and contribution of our framework, and your interest in our human simulation pipeline. This recognition is also kindly shared by other reviewers (Zf7z, irvc, PjNb). Following your suggestions, we conduct additional experiments to strengthen the analysis. For all HRI-related experiments, we report results from our main method under the more challenging collaboration type 2 (one trial per setting) due to time constraints. Below, we address each of your concerns.

Q1: Why is a fixed daily routine selected for the simulated human agent instead of enabling interaction to the robot?

A1: Thank you for pointing this out! We believe “fixed vs. dynamic” routines can be interpreted in two ways. The first is if a human with specific traits shows varied daily behavior while following a general routine (e.g., Monday 9am for cleaning, Tuesday 9am for exercise). We achieve this by setting a high LLM temperature (line 171). The second concerns if the human adjusts plans based on external factors such as mood, weather, or the robot’s behavior.

While our framework shows robustness to temporal factors by offline real human collaboration (Section 4.4, Table 3), we did not explicitly simulate human adaptation to robot behavior. Enabling reactive human behavior requires a cognitively richer agent and defining clear criteria for when and how adaptation occurs. This introduces additional complexity, turning the setup into a multi-agent framework. Our current goal is to develop an assistive robot that can improve across long-horizon, and we view simulating mutual adaptation as promising future work.

Yet, we fully agree it is valuable to explore human plan adaptation in response to robot behavior. We conduct two experiments to study this extension.

1. Increased human-robot interaction (Increased HRI): The human’s hourly intention/task is influenced by the robot’s action in the previous hour. For example, if the robot fails to clean the living room, the human may skip yoga to finish cleaning. We modify the human prompt to include robot's success/failure. We also ask the human to provide hourly feedback with reasoning to the robot (e.g., I disapprove this task because...). This helps the robot understand its mistakes and adjust its next action accordingly in the next hour.

2. Human-in-the-Loop (HITL): Three real humans replace the LLM and collaborate under setting 2. Users propose intentions based on their traits and adjust to robot predictions in real time. Users are shown retrieved object/motion sets and choose which to interact with and use. For simplicity and due to time constraints, users input responses as text instead of using keyboard to control the simulated agent.

The results are interesting. For Increased HRI, we expect higher or comparable performance due to frequent feedback but observe a slight drop. We find that the robot often repeats the previous intention to “make amends” when the human disapproves, while the simulated human sticks to its traits and only notes the error in feedback. For HITL, we expect lower performance due to real human variability, yet results are higher. After discussing with participants, we find they tend to recall and follow their mid-week routines, which are relatively stable. This consistency leads to lower variability than our simulated humans, allowing the robot to personalize more quickly and validating the behavioral diversity in our simulated humans.

Q2: What is the performance of the baseline(s) after sufficient days (in contrast to test-time learning setting)?

A2: Good question! To clarify, our framework includes two layers of test-time learning. The first layer operates within a day by prompting the robot-VLM with retrieved interaction history and the inferred human profile. The second layer operates across days by finetuning the robot-LLM classifier at the end of each day using data labeled from human feedback over previous days. For HRI at day t = n, the robot trains on data from days 0 to n–1, resulting in progressively more personalized behavior.

We choose this test-time learning setup for two reasons. First, within-day adaptation improves both success rate throughout a day and final success rate, as in Fig. 6a and Table 5. Second, across-day adaptation reflects a realistic deployment scenario: when a user purchases a home assistant or cleaning robot, they should be able to deploy it immediately, and the robot should improve over time through daily interactions. This contrasts with prior HRI setups that rely on collecting all data upfront and training the model only once [1, 2].

Yet, we fully agree it is valuable to study the train-all-at-once setting. We conduct two additional experiments. (1) We extend the benchmark by doubling the collaboration duration to observe continued adaptation trends. (2) we remove across-day learning and train the robot once after the final day, then evaluate its performance on all days.

For the test-time setting, performance stabilizes around day 8. While our benchmark uses 5-day collaboration, our main interest lies in the personalization trend, and longer HRC is easily supported. By day 8, the robot has learned most human preferences shaped by traits and routines, forming reliable alignment. Given the 7B model with LoRA finetuning, further gains become marginal without more capable models or finer-grained signals like multimodal cues or reasoning chains.

For the train-all-at-once case, daily performance is slightly higher or on par with the final days of test-time learning. This is because the robot remains unoptimized across all days, leading the simulated humans to disapprove more tasks. These disapprovals generate more negative samples for training. However, this comes at the cost of lacking any further personalization after the initial collaboration on day 1.

Q3: In Fig. 12, there is a pattern across days (increase → decrease within a day, higher increase on second day). Is this pattern due to your task setting or the learning paradigm of the agent?

A3: This is a great question! From Fig. 12, we observe performances tend to decrease around midday and recover toward the end of the day. We believe this pattern is due to two reasons.

First, it reflects how human routines are structured both in reality and in our simulation. Generally, humans tend to engage in more predictable activities in the morning and evening (e.g., hygiene, eating, relaxing). These behaviors are easier for the robot to predict and assist. In contrast, midday behavior tends to be more diverse and strongly aligned with human traits. Since our human simulation pipeline samples intentions/tasks based on both traits and time, this increases the diversity of midday intentions/tasks, making inference and alignment harder for the robot. As a result, performance temporarily drops. The performance recovers as human routines re-stabilize in the evening or next morning. Importantly, despite these fluctuations, performance improves overall across days.

Second, from Fig. 12, this fluctuation is amplified in more challenging task settings defined in our framework (Section 3.1). In Setting 1 (same human, same scene), the robot benefits from consistent exposure to the same human and environment, so midday dips are relatively small. In Setting 2 (same human, different scenes), unfamiliar object layouts introduce grounding challenges that increase midday variability. In Setting 3 (different humans, same scene), rotating between humans interrupts personalization, making intention interpretation harder, especially around midday when behavior is less routine. Finally, Setting 4 (different humans, different scenes) combines both factors, resulting in the largest fluctuations as the robot must generalize across both human and spatial contexts.

To further validate this pattern, we conduct two additional experiments by replacing the robot-VLM with a stronger model (GPT-4o) and a weaker one (Qwen2.5-VL-3B-Instruct). To quantify temporal fluctuation independently of overall performance trend, we compute the Coefficient of Variation of Differences (CVD), defined as the standard deviation of hour-to-hour differences normalized by the mean performance (in %).

Success Rate

CVD

As expected, stronger models produce better performance [3]. Notably, (1) the performance gap between GPT and llama is greater than between llama and Qwen. (2) while all VLMs exhibit fluctuation, stronger models like GPT show lower CVD. We attribute this to stronger models not only achieve better visual grounding for intention prediction but also generate more coherent intention-task proposals across time by better understanding prior interaction history.

[1] Watch-and-help: A challenge for social perception and human-ai collaboration. ICLR 2021.

[2] Partnr: A benchmark for planning and reasoning in embodied multi-agent tasks. ICLR 2025.

[3] Embodied Agent Interface: Benchmarking LLMs for Embodied Decision Making. NeurIPS 2024.