Smart Help: Strategic Opponent Modeling for Proactive and Adaptive Robot Assistance in Households

1. Introduction

• The challenge responds to assistive technologies that prioritize task completion while overlooking users’ self-esteem, self-efficacy, autonomy, and emotional acceptance of help.The motivation links physical, emotional, and aging-related difficulties to concerns about diminished self-worth and boundary violation.Source: §1 · p. 1
• Smart Help provides proactive, adaptive assistance that selectively addresses capability-based bottlenecks rather than taking over the entire household task.The Smart Helper infers goals and capability distributions, then assists only with goals the user cannot independently complete.Source: §1 · p. 1, Figure 1 · p. 2
• The authors build a realistic AI2-THOR 3D home environment for diverse capability distributions, dynamic goals, and general daily household tasks.The environment includes dimensions such as toggling, opening, closing, weight, and height, extending beyond specialized assistive-robot settings.Source: §1 · pp. 1–2
• The benchmark model jointly infers goals and capabilities, reasons about bottlenecks, and learns an online helping policy evaluated with six assistance-focused metrics.These metrics emphasize the helper’s contribution and the essentiality of assistance rather than only cumulative reward or task completion.Source: §1 · p. 2
• The paper reports that rigorous experiments validate the model components and that the holistic approach outperforms established baselines.The listed contributions include holistic evaluations supporting the benchmark model’s efficacy.Source: §1 · pp. 2–3

2. Related Work

• Assistive-robot research covers mobility, companionship, robotic arms, and socially assistive applications, while raising concerns about autonomy, privacy, acceptance, and trust.Many systems rely on simple rules or programs for social interaction.Source: §2 · p. 3
• Embodied collaboration tasks often target limited settings such as navigation or furniture rearrangement, whereas this work studies agents with different capabilities assisting vulnerable users in household tasks.Watch-and-Help [42] and NOPA [43] examine agents with the same capabilities, unlike the setting here.Source: §2 · p. 3
• Opponent-modeling methods have been applied to infer goals and anticipate behavior, but prior coordination work often overlooks the main agent’s capabilities.The paper addresses this gap by modeling capabilities and learning smart, adaptive, empathetic assistance under partial observation and uncertainty.Source: §2 · p. 3

3.1. Problem Formulation

• The helper’s objective is to assist the main agent’s goals while considering the main agent’s emotional state in a two-agent POMDP.The formulation includes physical and mental states, joint actions, local observations, transitions, shared rewards, and discounting.Source: §3.1 · p. 3
• The reward combines main-agent goal achievement with emotion-related reward, while omitting the helper’s own goal and setting β = 1.λe controls sensitivity to the main agent’s emotional state, and the helper exists solely to assist the main agent.Source: §3.1 · pp. 3–4
• Smart Help requires agents to determine not only how to assist, but also when assistance is needed and valued.This design distinguishes Smart Help from assistance tasks that do not explicitly model the timing and value of help.Source: §3.1 · p. 4

3.2. Challenge Implementation

• The AI2-THOR environment assigns main agents varied capabilities and hidden goals across household tasks, while the helper infers bottlenecks from observations.The environment includes symbolic and visual observations, but goals and capabilities remain hidden from the helper.Source: §3.2 · pp. 4–5
• Agents use intentional actions that specify target states and are decomposed by a low-level planner into executable simulator actions.The implementation defines seven intentional-action types and 373 possible intentional actions.Source: §3.2 · p. 4, Figure 2 · p. 5

4. Our Model

• The model addresses partial observation and hidden task information by estimating the main agent’s current goal and capabilities before selecting helper actions.Its policy factorizes into opponent modeling, P(gt, η|o0:t), and action selection conditioned on those predictions, P(a|gt, η, o0:t).Source: §4 · p. 5
• Training proceeds in two stages: opponent modeling is learned from collected trajectories, then the helping policy is trained with the pretrained model in the environment.The first stage supports anticipating main-agent behavior; the second handles dynamic and complex interactive tasks.Source: §4 · p. 5

4.1. The Opponent Modeling Module

• The opponent modeling module uses a sliding window because goals are transient, enabling more immediate inference as new actions arrive.The window examines recent trajectory segments and shifts over time; w denotes the window size.Source: §4.1 · pp. 5–6
• Object and agent encoders, a learnable object token, and MLP or transformer processing produce state, capability, and target-state features.Windowed state features are transformed into ξs, then a feature extractor derives ξη and ξt_g.Source: §4.1 · p. 6

4.2. The Helping Policy Module

• The helping policy uses an actor-critic architecture with an MLP and LSTM for action selection and context, plus a value MLP for state valuation.At each step it conditions the policy on the current observation and opponent-model features.Source: §4.2 · p. 6

5.1. Two-stage Learning

• The model first pretrains four auxiliary classifiers for capability and goal features, then freezes opponent modeling while training the helping policy with PPO across randomized AI2-THOR rooms, tasks, and capability distributions.Training uses 20 rooms and evaluates on 10 unseen rooms; progressive learning adds a warm-up phase followed by smart-help training.Source: §5.1 · p. 6

5.2. Reward

• The reward gives 20 for completing a main-agent goal, penalizes unnecessary help by -30, applies per-step costs, and assigns -20 when the scenario ends unfinished.Illegal actions receive a larger per-step cost of -0.5 instead of the normal -0.12.Source: §5.2 · p. 6

5.3. Baselines

• Baselines include random, end-to-end, MCTS, heuristic MCTS, oracle-goal planners, and an LLM helper evaluated from AI2-THOR observations.The oracle-goal variants use either the true main-agent goal or a random goal, while the LLM is prompted at every step.Source: §5.3 · pp. 6–7

5.4. Ablation Study

• Ablations remove capability embeddings, alter the emotional-reward coefficient, or replace progressive learning to isolate the contributions of opponent modeling and training strategy.The full model uses progressive learning with λe = 1.0 after warming up basic helper skills.Source: §5.4 · p. 7

5.5. Evaluation Metrics

• Evaluation combines task and goal success, Helping Necessity, Helping Rate, Episode Length, SPL, and average reward, targeting necessary assistance without sacrificing completion or efficiency.Helping Necessity measures whether assistance is required by the main agent’s capability, while Helping Rate measures helping when help is needed.Source: §5.5 · p. 7

Qualitative Results

• The full model helps only with the failed bottleneck “(Open, Microwave)” and lets the main agent complete remaining goals independently, unlike BaseModel-PL-λe=0.0.The baseline also helps with “(ToggleOff, Microwave)” after addressing the bottleneck.Source: Figure 3 · p. 8, §6 · p. 8

Comparison with baselines

• Higher SR, GSR, HN, and HR, with similar EL and SPL, distinguish the full model from non-oracle baselines; MCTSTG remains the upper bound with true goals.The LLM has the shortest EL but underperforms on other metrics, while MCTS-heuristic outperforms pure MCTS.Source: §6 · pp. 7–8, Table 1 · p. 8

Ablation analysis

• After warm-up, setting λe = 1.0 in progressive learning greatly improves Helping Necessity while maintaining competitive performance across other metrics.The warm-up phase encourages basic task completion, whereas the second phase focuses on assistance based on the main agent’s needs and feelings.Source: §6 · p. 8

7. Conclusion

• The work contributes a Smart Help challenge, an environment, and a model aimed at fostering smarter and more harmonious human-artificial-agent interaction.The authors present these resources as foundations for future studies of the problem.Source: §7 · p. 8

Supplementary Material

• Task parsing converts overarching activities into specific goals, but the resulting sequence can vary across scenes and probabilistically within a scene.For MakeCoffee, a mug already in the coffee machine reduces the required goals to activating and deactivating the machine.Source: Table 1 · p. 12

• Appendix A provides the detailed environment setting.Source: Overview · p. 12
• Appendix B provides the model implementation details.Source: Overview · p. 12
• Together, the appendices document both the task environment and the model implementation.Source: Overview · p. 12

• 30 AI2-THOR kitchen scenes are split into 20 training scenes and 10 testing scenes, using only objects common across scenes.Source: A. Environment Setting · p. 12, Figure 1 · p. 13
• The tasks are MakeBreakfast, ArrangeRoom, and MakeCoffee, each parsed into room-dependent goal sequences.Source: A. Environment Setting · p. 12
• Goals specify target states and may require ordered execution and maintaining achieved states until subsequent goals are completed.Source: A. Environment Setting · p. 12
• Seven intentional actions are translated by a low-level planner into executable simulator actions whose validity depends on context.Source: A. Environment Setting · p. 12
• Agent capabilities represent reachable height, liftable weight, and abilities to open, close, toggle on, and toggle off objects.Source: A. Environment Setting · pp. 12–13

B.1. Trajectory dataset

• The trajectory dataset contains 60,024 trajectories with five observation time points and the main agent’s final goal, using expert main-agent behavior and randomized helper actions.Source: §B.1 · pp. 13–14

B.2. Opponent Modeling Module

• The opponent modeling module jointly represents five-step observations to predict the main agent’s goal, relevant objects, and capabilities.Source: §B.2 · pp. 14–15
• The opponent model achieves 83.6% prediction accuracy when inferring human-agent goals from constrained observations.Source: §B.2 · p. 15

B.3. Helping Model

• The helping model concatenates latent opponent representations with observations before using a linear layer, ReLU, and 512-unit LSTM to generate actions.Source: §B.3 · p. 15

B.4. LLM

• The LLM planner receives environment and partial agent observations, outputs helper actions, and replaces invalid outputs with Wait.Source: §B.4 · p. 15, Figure 3 · p. 16, Figure 4 · p. 16

B.5. MCTS-heuristic

• MCTS evaluates rollouts up to five steps and returns the most visited action after 500 simulations, while heuristic sampling interpolates between random exploration and rule-based sequences.Source: §B.5 · pp. 15–16
• MCTSTG achieves the best performance across all metrics, but its Helping Necessity is relatively low because it uses a simple takeover strategy.Source: §B.5 · p. 17