Research map / Probabilistic and scientific modeling / Molecules and proteins
Smart Help: Strategic Opponent Modeling for Proactive and Adaptive Robot Assistance in Households
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the Smart Help task in an AI2-THOR household simulator, where a helper infers a main agent's goals and capabilities via an opponent-modeling module and learns adaptive, minimal assistance with PPO.
Problem
Assistive robots for vulnerable users typically take over tasks without considering the user's capabilities, comfort, and self-efficacy; existing cooperative embodied tasks ignore the helped agent's differing capabilities.
Approach
Jointly infer the main agent's goal and capability distribution from partial observations (sliding-window opponent model), then train a helping policy with an emotion-aware reward that penalizes helping with goals the user can do alone, so the helper gives only bottleneck assistance.
Key results
- Full model (BaseModel-PL-λe=1.0) reaches HN 0.498 and HR 0.506, versus 0.345/0.381 for the LLM baseline and 0.267/0.315 for End2end.
- MCTSTG, which knows the true main-agent goal, achieves the best SR of 0.593 and GSR 0.804, serving as an oracle upper bound.
- Full model achieves SR 0.483 and SPL 0.200, with reward -10.625, the best among non-oracle methods on reward.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use opponent modeling instead of end-to-end action generation | Smart Help helper · Smart Help evaluation | SR: 0.267 -> 0.419 | Helped |
| Add capability embedding to helping policy training (ablation) | BaseModel · Smart Help evaluation | SR: 0.374 -> 0.419 | Helped |
| Set λe=1.0 during the 840-epoch training phase (ablation) | BaseModel · Smart Help evaluation | SR: 0.419 -> 0.181 | Hurt |
| Set λe=1.0 during the 840-epoch training phase (ablation) | BaseModel · Smart Help evaluation | HN: 0.408 -> 0.193 | Hurt |
| Use progressive learning with λe=0.0 in the second phase (ablation) | BaseModel · Smart Help evaluation | SR: 0.419 -> 0.488 | Helped |
| Use progressive learning with λe=0.0 in the second phase (ablation) | BaseModel · Smart Help evaluation | HN: 0.408 -> 0.455 | Helped |
| Set λe=1.0 in the second phase after warm-up (ablation) | BaseModel-PL-λe=0.0 · Smart Help evaluation | HN: 0.455 -> 0.498 | Helped |
| Set λe=1.0 in the second phase after warm-up (ablation) | BaseModel-PL-λe=0.0 · Smart Help evaluation | SR: 0.488 -> 0.483 | No clear effect |
| Use the full Smart Help method instead of MCTSTG | Smart Help helper · Smart Help evaluation | HN: 0.610 -> 0.498 | Hurt |
Surprising findings
- The LLM helper has the shortest episode length (22.9) but lower success and helping necessity than the proposed model.
- Setting the emotion term λe=1.0 from scratch (BaseModel-λe=1.0) performs worse than λe=0.0, so the warm-up stage is needed before adding the emotional reward.
Limitations
- The main agent's emotional response is modeled only as a reward penalty, not measured with real users.
- Evaluation uses simulated agents and a limited set of three household tasks and 10 test rooms.
Experimental setup
Datasets: AI2-THOR simulated trajectories
Benchmarks: Smart Help (AI2-THOR, 10 held-out rooms, 14 task-capability pairs)
Metrics: SR GSR HN HR EL SPL average reward
Baselines: Random End2end LLM (gpt-3.5-turbo-instruct) MCTS MCTS-heuristic MCTSTG MCTSRG
Builds on: Watch-and-Help NOPA PPO MCTS
Closest papers by meaning
- LIBERO-PeRM: Benchmarking Personalized Robotic Manipulation (NeurIPS 2026 · similarity 0.80)
- Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance (ICLR 2025 · similarity 0.80)
- PARTNR: A Benchmark for Planning and Reasoning in Embodied Multi-agent Tasks (ICLR 2025 · similarity 0.79)
- Promptable Behaviors: Personalizing Multi-Objective Rewards from Human Preferences (CVPR 2024 · similarity 0.79)
- Learning to Assist Humans without Inferring Rewards (NeurIPS 2024 · similarity 0.79)
- ProAct: A Benchmark and Multimodal Framework for Structure-Aware Proactive Response (ICML 2026 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.