Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Back to the Future of Integrated Robot Systems
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.
The paper argues that robot architectures should be built on classical AI principles (refinement, ecological rationality, interactive learning) with data-driven methods as one tool among many, and illustrates this with the author's prior work.
Problem
Deep, data-driven robotics methods are resource-hungry, opaque, and make arbitrary decisions in novel situations, while practical robots need transparent multi-step, multi-level decision making and ad hoc collaboration under resource constraints and open-world uncertainty.
Approach
Revisit early-AI principles, namely refinement across abstractions, ecological rationality with simple decision heuristics, and interactive learning triggered as needed, and embed them in robot architectures so that deep networks become one tool within a principled reasoning-and-learning foundation.
Key results
- Ad hoc teamwork predictive models based on FF-tree ensembles used about 5K training examples instead of 1M for state-of-the-art methods, with comparable or better performance.
- Refinement-based architecture for vision-based scene understanding achieved performance comparable with or better than deep-network-only systems while using orders of magnitude fewer training examples (as reported in cited prior work).
- LLM-plus-decision-heuristics-plus-classical-planning for home task collaboration showed substantial gains in accuracy and computational efficiency over LLM-only systems (cited prior work).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a refinement-based architecture with prior domain knowledge and rapidly revised models built from an ensemble of fast-and-frugal trees | Ad hoc teamwork agent · Ad hoc teamwork in complex environments with previously unknown changes | collaboration performance | Helped |
| Use an ensemble of fast-and-frugal trees for predictive models | Ad hoc teamwork agent · Ad hoc teamwork | training examples: 5K instead of 1M | Helped |
| Use reasoning and learning together, with deep networks only as needed and on relevant examples | Visual scene understanding and planning architecture · Visual scene understanding, planning, and question answering | performance | Helped |
| Use deep networks only as needed and with relevant examples | Visual scene understanding and planning architecture · Visual scene understanding, planning, and question answering | training speed | Helped |
| Bootstrap reasoning and learning off each other | Visual scene understanding and planning architecture · Visual scene understanding, planning, and question answering | performance over time | Helped |
| Anticipate future tasks using an LLM, then plan jointly for anticipated and current tasks | Embodied AI agent collaborating with a human · Home task completion | accuracy | Helped |
| Anticipate future tasks using an LLM, then plan jointly for anticipated and current tasks | Embodied AI agent collaborating with a human · Home task completion | computational efficiency | Helped |
| Remove future-task anticipation (ablation) | Embodied AI agent collaborating with a human · Home task completion | accuracy | Hurt |
Surprising findings
- The author argues that many papers incorrectly claim LLMs can plan, citing a position paper (Kambhampati et al. 2024).
Limitations
- Argument is largely position-based; the architectural examples are summarized from prior work without new experiments in this paper.
- Refinement-based approaches do not yet fully handle bidirectional information flow, uncertainty from non-stationarity, or contact discontinuities, as the author notes for existing work.
Experimental setup
Baselines: deep network systems data-driven ad hoc teamwork methods
Builds on: neurosymbolic AI bounded rationality fast and frugal trees LLM-Modulo
Closest papers by meaning
- Deep Reinforcement Learning for Robotics: A Survey of Real-World Successes (AAAI 2025 · similarity 0.85)
- Zero-Shot Robotic Manipulation with Pre-Trained Image-Editing Diffusion Models (ICLR 2024 · similarity 0.84)
- ManiTaskGen: A Comprehensive Task Generator for Benchmarking and Improving Vision-Language Agents on Embodied Decision-Making (NeurIPS 2026 · similarity 0.84)
- Reasoning in Visual Navigation of End-to-end Trained Agents: A Dynamical Systems Approach (CVPR 2025 · similarity 0.84)
- What Matters in Learning from Large-Scale Datasets for Robot Manipulation (ICLR 2025 · similarity 0.84)
- DiffusionVLA: Scaling Robot Foundation Models via Unified Diffusion and Autoregression (ICML 2025 · similarity 0.84)
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.