Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Structured agents for physical construction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use relative object-centric actions instead of absolute actions | Construction agents · Construction tasks | reward: up to 1.7 times more reward; factors of up to 2.4 | Helped |
| Add MCTS with train budget 10 and test budget 100 | GN-DQN · Reaching, hardest scenes | median reward: 2.17 to 2.72 | Helped |
| Add MCTS with train and test budgets of 10 | GN-DQN · Coveringone | median reward: 3.60 to 4.61 | Helped |
| Increase recurrent propagation steps from 1 to 3 (ablation) | GN-DQN · Silhouette | median reward: 3.75 to 4.04 | Helped |
| Increase recurrent propagation steps from 3 to 5 (ablation) | GN-DQN · Silhouette | median reward: 4.04 to 4.06 | Helped |
| Increase recurrent propagation steps from 1 to 3 (ablation) | GN-DQN · Reaching | median reward: 2.49 to 2.84 | Helped |
| Increase recurrent propagation steps from 3 to 5 (ablation) | GN-DQN · Reaching | median reward: 2.84 to 2.81 | Hurt |
| Increase recurrent propagation steps from 1 to 3 (ablation) | GN-DQN · Coveringtwo | median reward: 3.41 to 3.96 | Helped |
| Increase recurrent propagation steps from 3 to 5 (ablation) | GN-DQN · Coveringtwo | median reward: 3.96 to 4.01 | Helped |
| Increase recurrent propagation steps from 1 to 3 (ablation) | GN-DQN · Coveringone | median reward: 2.62 to 3.03 | Helped |
| Increase recurrent propagation steps from 3 to 5 (ablation) | GN-DQN · Coveringone | median reward: 3.03 to 3.02 | Hurt |
| Use a fully connected rather than sparse input graph (ablation) | GN-DQN · Construction tasks | performance | Helped |
Closest papers by meaning
- Blocks Assemble! Learning to Assemble with Large-Scale Structured Reinforcement Learning (ICML 2022 · similarity 0.90)
- I-PHYRE: Interactive Physical Reasoning (ICLR 2024 · similarity 0.84)
- Entity-Centric Reinforcement Learning for Object Manipulation from Pixels (ICLR 2024 · similarity 0.83)
- Reasoning in Visual Navigation of End-to-end Trained Agents: A Dynamical Systems Approach (CVPR 2025 · similarity 0.83)
- PhyBlock: A Progressive Benchmark for Physical Understanding and Planning via 3D Block Assembly (NeurIPS 2025 · similarity 0.83)
- Kinetix: Investigating the Training of General Agents through Open-Ended Physics-Based Control Tasks (ICLR 2025 · similarity 0.83)
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.