Skip to content

Embodied-R1: Reinforced Embodied Reasoning for General Robotic Manipulation

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

ICLR 2026; Embodied-R1: unified pointing representation for generalist manipulation

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Embodied-R1 is a 3B VLM trained via Reinforced Fine-tuning (RFT) that uses "pointing" as a unified, embodiment-agnostic intermediate representation to bridge visual perception and robotic action. It defines four grounding/trajectory abilities, builds a 200K-sample dataset, and achieves 87.5% zero-shot success on 8 real-world XArm tasks without task-specific fine-tuning.

Problem

Existing VLA models fail to generalize to novel settings due to (a) data scarcity—limited embodied data prevents grounding language/vision with physical actions—and (b) embodiment heterogeneity—diverse robot morphologies obstruct knowledge transfer. Prior pointing-based approaches are either too narrow (single signal type) or rely on rigid CoT templates via SFT that cannot generalize, and SFT struggles with the multi-solution dilemma inherent in spatial tasks (many valid pointing answers for the same instruction).

Method

Embodied-R1 is built on Qwen2.5-VL-3B and defines four embodied pointing abilities: 1. REG (Referring Expression Grounding): localizes an object via a point \(p=(p,q)\) inside its segmentation mask. 2. RRG (Region Reference Grounding): identifies a free-space placement region from relational language. 3. OFG (Object Functional Grounding): marks functionally critical affordance points (e.g., grasp handles). 4. VTG (Visual Trace Generation): produces an ordered point sequence \(\tau = \{p_t\}_{t=1}^{T}\) as an object-centric manipulation trajectory.

Training is a two-stage RFT curriculum using GRPO: - Stage 1: Spatial reasoning on Embodied-Spatial-84K (from SAT + WhatsUp) + ViRL-Subset-18K (general knowledge to prevent catastrophic forgetting). - Stage 2: Embodied pointing on Embodied-Points-200K, structured as "question–verification" pairs (not question–answer), enabling RFT to reward any geometrically correct solution.

A modular reward library combines: format reward \(r_\text{format}\), accuracy reward \(r_\text{acc}\), point-in-mask reward \(r_\text{mask}\), dense Euclidean distance reward \(r_\text{dis}\), and trajectory RMSE reward \(r_\text{trace}\). Per-task total reward \(R = \sum_{r \in F} w_r \cdot r\) is normalized to \([0,1]\) with task-specific weights.

At inference, outputs route to either (a) an Affordance Points Branch feeding CuRobo motion planner, or (b) a Visual Trace Branch that lifts the 2D trace to 3D via pinhole projection and executes in SE(3).

Key Contributions

  • Establishes "pointing" as a unified, embodiment-agnostic intermediate representation with four formalized abilities (REG, RRG, OFG, VTG).
  • Releases Embodied-Points-200K, a curated dataset using "question–verification" pairs to support RFT training across all four pointing tasks.
  • Proposes Embodied-R1, a 3B RFT-trained VLM with a multi-task reward system that resolves the multi-solution dilemma inherent in spatial grounding.
  • Demonstrates zero-shot cross-embodiment generalization without task-specific fine-tuning on both simulation and real hardware.

Results

  • Spatial reasoning (11 benchmarks, avg. rank): Embodied-R1 achieves mean rank 2.1 among open-source models, vs. 3.4 (Embodied-R1 w/o CS), 3.7 (Embodied-SFT), 4.4 (RoboBrain-7B), 4.6 (FSD-13B).
  • Pointing benchmarks:
  • RoboRefit: 85.58% (vs. FSD 56.73%, RoboPoint 49.82%)
  • Where2Place: 69.50% (vs. FSD 45.81%, RoboPoint 46.01%)
  • VABench-P: 66.00% (vs. FSD 61.82%, Qwen2.5VL 9.89%)
  • Part-Afford: 56.63% (vs. RoboPoint 27.60%, Qwen2.5VL 23.42%)
  • VABench-V (trajectory): RMSE 77.8, MAE 45.0, LLM Score 7.3 (vs. FSD: 78.3 / 63.4 / 6.2).
  • SIMPLEREnv (simulation): 56.2% success rate zero-shot.
  • Real-world XArm (8 tasks): 87.5% zero-shot success rate, representing a 62% improvement over strong baselines.

Limitations

  • The pointing-to-action pipeline requires a separate motion planner (CuRobo) and depth sensor; errors in depth estimation or planner failures are not addressed by the VLM.
  • The 3B model, while competitive, still lags closed-source models (GPT-4o) on several spatial benchmarks.
  • VTG data construction relies on multiple pre-trained vision models (GPT-4o, CoTracker3, Grounded-SAM), introducing noise that requires iterative rule-based filtering.
  • Experiments are limited to a single robot platform (XArm); cross-embodiment transfer to morphologically distinct platforms (bipeds, dexterous hands) is not validated.
  • The action executor pipelines are open-loop for the Visual Trace Branch; closed-loop correction is only mentioned as future work via integration with learning-based methods.

Relevance to Vision-Language Models

Embodied-R1 directly advances the VLM-for-robotics line of work by demonstrating that RFT with verifiable spatial rewards produces qualitatively stronger out-of-distribution generalization than SFT-based CoT approaches, a finding that parallels gains seen in mathematical and code reasoning VLMs. The paper operationalizes pointing as a compact, verifiable output space for VLMs—analogous to bounding boxes in detection, but richer—enabling reward-based training without environment rollouts. For researchers tracking VLMs, the key insight is that framing embodied grounding as "question–verification" rather than "question–answer" unlocks RL-style training on internet-scale visual data, directly scaling VLM spatial understanding to physical control. The modular reward design and two-stage curriculum are immediately transferable to other spatial VLM tasks beyond manipulation.