Research map / Multi-agent systems and games / Human-AI and goal-conditioned
GenPlan: Generative Sequence Models as Adaptive Planners
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.
Proposes GenPlan, an energy-guided discrete-flow-matching (CTMC denoising) sequence planner that jointly generates goals, states and actions to adapt from single-goal demonstrations to multi-task, unseen BabyAI missions.
Problem
Sequence-model planners trained on demonstrations of single-goal tasks fail to adapt to multi-task, order-critical missions with unseen goals, doors, and obstacles, and to generalize to new environments with walls and locked doors.
Approach
Frames planning as iterative denoising of discrete goal, state and action sequences via discrete flow matching with a CTMC sampler, trained with an energy (negative pseudo-likelihood) objective and an entropy lower bound that encourages stochastic exploration and task discovery.
Key results
- On BabyAI adaptive planning (AP) tasks, GenPlan-M reaches 52.3% mean success versus 38.4% for LEAP (conditional) and 18.2% for DT, a gain of over 10 points.
- On BabyAI trajectory planning, GenPlan-M achieves 41.5% mean success vs 31% for LEAP and 33.9% for DT.
- On Franka Kitchen, GenPlan-M completes 3.40 subtasks vs 2.62 for DP-C, while VQ-BeT leads at 3.66.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GenPlan-M instead of DT | BabyAI adaptive planning · BabyAI AP mean | success rate: 52.3% vs 18.2% | Helped |
| use GenPlan-M instead of VQ-BeT | Kitchen · Kitchen | # Tasks: 3.40 vs 3.66 | Hurt |
| add joint goal and state prediction (ablation) | GenPlan-M · MazeS4G2 (TP) | success rate: 34.4% vs 32%; ↓ 7.6% | Helped |
| add joint goal and state prediction (ablation) | GenPlan-M · KeyCorS3R3 (IC) | success rate: 13.6%; ↓ 4% | Helped |
| add joint goal and state prediction (ablation) | GenPlan-M · DoorsOrder (IC) | success rate: 15.6%; ↓ 19.6% | Helped |
| add joint goal and state prediction (ablation) | GenPlan-M · SeqS5R2Un (AP) | success rate: 32.8%; ↓ 9.2% | Helped |
| add entropy regularization (ablation) | GenPlan-M · MazeS4G2 (TP) | success rate: 34.4%; ↓ 5.2% | Helped |
| add entropy regularization (ablation) | GenPlan-M · KeyCorS3R3 (IC) | success rate: 7.6%; ↓ 10% | Helped |
| add entropy regularization (ablation) | GenPlan-M · DoorsOrder (IC) | success rate: 34.4%; ↓ 0.8% | Helped |
| add entropy regularization (ablation) | GenPlan-M · SeqS5R2Un (AP) | success rate: 34.8%; ↓ 7.2% | Helped |
| retain history context (ablation) | GenPlan-M · KeyCorS3R3 (IC) | success rate: 0.4%; ↓ 17.2% | Helped |
| retain instruction embeddings (ablation) | GenPlan · Doors Order (IC) | success rate: 22%; 13.2% ↓ | Helped |
Surprising findings
- Energy-gradient guided DDPM sampling gets trapped in local minima, repeating frequent dataset actions such as forward moves (0-2% success in some tasks).
- Entropy regularization hurts instruction-completion tasks requiring open or pickup actions, while helping adaptive planning.
Limitations
- The entropy lower bound beta is a manually tuned hyperparameter.
- Assumes near-optimal demonstrations; robustness to suboptimal data only partly studied.
- Language grounding is left for future work and instruction ablation shows dependence on instruction embeddings.
Experimental setup
Datasets: BabyAI demonstrations (single-goal tasks) PushT human demonstrations Franka Kitchen demonstrations MetaWorld tasks
Benchmarks: BabyAI PushT Franka Kitchen MetaWorld
Metrics: success rate IoU coverage number of tasks completed
Baselines: LEAP LEAP without goal conditioning Decision Transformer Diffusion Policy (CNN) Diffusion Policy (Transformer) VQ-BeT MDDT
Builds on: Decision Transformer LEAP discrete flow models Planning Transformer VQ-BeT
Closest papers by meaning
- System 1.x: Learning to Balance Fast and Slow Planning with Language Models (ICLR 2025 · similarity 0.88)
- Behavioral Exploration: Learning to Explore via In-Context Adaptation (ICML 2025 · similarity 0.87)
- GenSim: Generating Robotic Simulation Tasks via Large Language Models (ICLR 2024 · similarity 0.87)
- GenPO: Generative Diffusion Models Meet On-Policy Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks (ICML 2025 · similarity 0.86)
- Diffusion Subgoal Planning for Long-Horizon Offline Goal-Conditioned Reinforcement Learning (NeurIPS 2026 · similarity 0.86)
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.