Research map / Vision-language and multimodal / Multimodal LLMs

Affordance-R1: Reinforcement Learning for Generalizable Affordance Reasoning in Multimodal Large Language Models

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 1 citations (OpenAlex)

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.

Affordance-R1 applies GRPO-based reinforcement learning with format, perception, and recognition rewards to an MLLM (Qwen2.5-VL-7B) for reasoning-based affordance grounding, paired with a new GPT-4o-rewritten dataset ReasonAff.

Problem

Affordance grounding requires reasoning about implicit instructions to locate functional object regions, but existing affordance models lack chain-of-thought reasoning and generalize poorly out of domain.

Approach

Train an MLLM purely with GRPO (no SFT on reasoning data) using a reward combining think/rethink/answer format checks, box IoU/L1/box-count perception rewards, and a word2vec affordance-recognition reward, then pass predicted boxes and points to SAM2 for masks.

Key results

• On ReasonAff, Affordance-R1 reaches 67.41 gIoU and 62.72 cIoU versus 63.04 gIoU for Vision Reasoner (best baseline).
• Zero-shot on UMD, Affordance-R1 gets 49.85 gIoU vs 44.26 for Seg-Zero.
• Zero-shot on AGD20K (Seen split), Affordance-R1 gets 31.78 gIoU vs 26.99 for Seg-Zero.

Surprising findings

• The paper reports that Affordance-R1 trained only with RL shows emergent test-time reasoning and robust VQA ability without VQA training data.

Limitations

• Inconsistent reported backbone and hyperparameters (appendix says Qwen-2-VL-7B and lr 1e-4, KL 0.05, versus main text Qwen2.5-VL-7B, lr 1e-6, KL 5e-3).
• Recognition reward relies on word2vec similarity threshold, a coarse semantic signal.
• Evaluation is on ReasonAff, UMD, and AGD20K only; the dataset is derived from Instruct-Part, so in-domain and source-dataset generalization is entangled.

Experimental setup

Datasets: ReasonAff Instruct-Part UMD Part Affordance AGD20K

Benchmarks: ReasonAff UMD AGD20K

Metrics: gIoU cIoU P@50 P@50:95 KLD SIM NSS

Baselines: VLPart OVSeg SAN LISA-7B SAM4MLLM AffordanceLLM Qwen2.5-VL-7B InternVL3-8B Seg-Zero Vision Reasoner

Builds on: GRPO Seg-Zero DeepSeek-R1 Instruct-Part

Closest papers by meaning

• AffordDP: Generalizable Diffusion Policy with Transferable Affordance (CVPR 2025 · similarity 0.89)
• AffordBot: 3D Fine-grained Embodied Reasoning via Multimodal Large Language Models (NeurIPS 2025 · similarity 0.89)
• SeqAfford: Sequential 3D Affordance Reasoning via Multimodal Large Language Model (CVPR 2025 · similarity 0.89)
• Robot-R1: Reinforcement Learning for Enhanced Embodied Reasoning in Robotics (NeurIPS 2025 · similarity 0.88)
• LaST-R1: Reinforcing Robotic Manipulation via Adaptive Physical Latent Reasoning (NeurIPS 2026 · similarity 0.88)
• GREAT: Geometry-Intention Collaborative Inference for Open-Vocabulary 3D Object Affordance Grounding (CVPR 2025 · similarity 0.87)