Research map / Vision-language and multimodal / Robot action and navigation

Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Embodied multimodal LLM agents · 0 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.

Embodied-R1.5 is an 8B embodied foundation model unifying cognition, planning/correction, and pointing via three automated data pipelines (15B tokens), SFT plus a multi-task balanced GRPO-style RL recipe, a PGC closed-loop framework, and a light VLA extension.

Problem

Existing embodied models cover only parts of physical intelligence (cognition, planning/correction, pointing) and suffer multi-task interference; they also lack closed-loop validation on long-horizon real-world tasks.

Approach

Unify spatial cognition, planning/correction, and pointing in one 8B VLM trained with plain-text coordinate outputs, a two-stage SFT+RFT recipe with batch-level std advantage normalization and difficulty filtering, and a single-model Planner-Grounder-Corrector loop; attach a flow-matching action expert for VLA without large action pretraining.

Key results

• Embodied-R1.5 (8B) achieves SOTA on 16 of 24 embodied VLM benchmarks with a 70.4% average over 21 main benchmarks, surpassing Gemini-Robotics-ER-1.5 by 17.0 points and GPT-5.4 by 21.7 points.
• Embodied-R1.5-VLA reaches 92.4% on SimplerEnv Google Robot Visual Matching, versus 72.7% for pi0.5, and 97.3% on LIBERO without action pretraining.
• On ManiSkill-Affordance, Embodied-R1.5 scores 0.76 on seen and 0.66 on unseen categories, versus ManipLLM's 0.59 and 0.54.

Surprising findings

• Zero-shot on RoboTwin2.0 (no RoboTwin data), Embodied-R1.5 reaches 65.0% average, close to fine-tuned pi0.5 (72.9%).
• Embodied-R1.5 without action pretraining matches or exceeds action-pretrained OpenVLA-OFT on LIBERO (97.3% vs 97.1%).

Limitations

• Operates only on 2D image inputs; lacks native 3D perception such as point clouds or depth.
• PGC closed-loop validated only on tabletop manipulation, not mobile manipulation or navigation.
• VLA extension uses a lightweight action head; coupling between reasoning and action is not explored.

Experimental setup

Datasets: VLM-3R Cambrian-S SAT DepthLM Robo2VLM Fractal BridgeData V2 DROID RoboVQA EgoPlan-IT RoboFAC RoboFail

Benchmarks: RoboVQA EgoPlan-2 Cosmos-Reason RoboFAC VABench-P Where2Place RefSpatial Part-Afford RoboRefit RoboAfford PointBench PIO

Metrics: accuracy success rate average nearest-neighbor point distance RMSE DFD

Baselines: Gemini-Robotics-ER-1.5 GPT-5.4 GPT-4o Gemini-2.5-Pro Qwen3-VL-8B RoboBrain-2.0-7B Embodied-R1 pi0.5 OpenVLA-OFT ManipLLM

Builds on: GRPO Embodied-R1 Qwen3-VL EasyR1 starVLA

Closest papers by meaning

• EmbodiedBench: Comprehensive Benchmarking Multi-modal Large Language Models for Vision-Driven Embodied Agents (ICML 2025 · similarity 0.89)
• OmniEVA: Embodied Versatile Planner via Task-Adaptive 3D-Grounded and Embodiment-aware Reasoning (ICLR 2026 · similarity 0.89)
• Towards Reliable Code-as-Policies: A Neuro-Symbolic Framework for Embodied Task Planning (NeurIPS 2025 · similarity 0.89)
• Embodied Agent Interface: Benchmarking LLMs for Embodied Decision Making (NeurIPS 2024 · similarity 0.89)
• SEEA-R1: Tree-Structured Reinforcement Fine-Tuning for Self-Evolving Embodied Agents (NeurIPS 2025 · similarity 0.89)
• Universal Actions for Enhanced Embodied Foundation Models (CVPR 2025 · similarity 0.88)