InternVLA-M1: A Spatially Guided Vision-Language-Action Framework for Generalist Robot Policy

• Robot policies must bridge indirect textual spatial cues to continuous embodied actions, which are scarcely represented in VLM training data; teleoperated datasets remain modest in scale and diversity.The introduction motivates embodiment-agnostic spatial priors as a bridge between instructions and embodiment-specific motor commands.Source: §1 · p. 1
• Hierarchical systems often depend on rule-based task decomposition and manual planning heuristics, while data-driven VLAs can overfit fine-grained motor behaviors and under-generalize to spatially demanding instructions.Transferable priors include object recognition, affordance grounding, visual trajectory reasoning, relative localization, and scaling; downstream learning can then specialize to embodiment-specific control.Source: §1 · pp. 1–2
• InternVLA-M1 combines a spatial-prior VLM planner with an action expert, using spatial grounding pre-training and spatially guided action post-training for embodiment-specific control.The framework was trained on over 3M multimodal samples, including 2.3M spatial grounding samples and 0.7M multimodal understanding samples; grounding supervision covers points, boxes, and traces.Source: §1 · pp. 2–3
• +5.9% and +9.8% average success-rate gains over its variant on SimplerEnv Google Robot and WidowX, respectively; 6.2% average improvement over prior works across 200 tabletop scenarios; and +20.6% success rate on unseen objects and novel setups.The tabletop result is under few-demonstration fine-tuning. In real-world clustered pick-and-place, the model also outperforms baselines by large margins in long-horizon tasks under perturbations such as physical interference and task replanning.Source: §1 · p. 3

2.1. Model Architecture

• InternVLA-M1 pairs a Qwen2.5-VL-3B-instruct spatial-prior System 2 with a diffusion-policy System 1 action expert in an approximately 4.1B-parameter model.The action expert uses DINOv2 and a lightweight state encoder; inference runs on one RTX 4090 with around 12 GB memory, and the VLM reaches approximately 10 FPS with FlashAttention.Source: §2.1 · p. 3, §2 · p. 3
• Joint multimodal and action supervision aggregates grounding and robot-demonstration losses in one update so perception and control co-adapt rather than learn in isolation.The VLM learns from real and synthetic spatial-grounding data, while the Action Expert specializes those priors into embodiment-specific motor commands.Source: §2.1 · pp. 3–4
• Spatial prompting cues the planner to reason about relevant objects and spatial constraints, while a lightweight querying transformer maps variable-length planner features to fixed query tokens.A gradient decay factor, such as 0.5, attenuates action-to-VLM gradients to preserve semantic reasoning during joint optimization.Source: §2.1 · p. 4

2.2. Training Recipe

• Stage 1 trains the VLM for robotics-focused spatial reasoning using web-scale multimodal corpora and robot datasets reformatted for grounding and reasoning tasks.The unified QA-style data covers bounding-box detection, trajectory prediction, affordance recognition, and chain-of-thought reasoning.Source: §2.2 · pp. 4–5
• Stage 2 jointly optimizes the VLM and Action Expert, alternating robot trajectories with spatial-grounding data to connect semantic understanding with motion generation.Trajectory batches update both components with an L2 noise-prediction loss, while grounding batches update only the VLM through next-token prediction.Source: §2.2 · p. 5

3.1. Spatial Grounding Data for Pre-training

• Over 3M multimodal pre-training samples include more than 2.3M spatial-reasoning examples across General, Box, Trajectory, and Point QA.The categories contain approximately 637K, 879K, 684K, and 832K samples, respectively. The datasets include diverse multimodal QA, grounding annotations, end-effector trajectories, and point-localization data.Source: §3.1 · pp. 5–6

3.2. Synthetic Data For Action Post-Pre-training

• 244K validated closed-loop pick-and-place samples form the synthetic InternData M1 dataset for action post-pre-training.The data pipeline uses the same object set and positional distributions as InternVLA-M1-Interface Data, with randomized lighting and textures to increase visual diversity.Source: §3.2 · p. 6

3.3. Scalable Synthetic Data Engine for Instruction-Following

• The scalable synthesis engine generates physically feasible, task-complete trajectories and decouples planning from rendering to create VLM and VLA data.It replays structured scene and trajectory records under randomized visual conditions, calibrates cameras to match real-world viewpoint geometry, and provides bounding boxes and end-effector trajectories. Its asset library contains 14K annotated objects, 211 tables, 1.6K textures, and 87 dome lights.Source: §3.3 · p. 7

4.1. Experiments on Public Benchmarks

• Public-benchmark evaluation covers SimplerEnv’s Google Robot and WidowX tasks and LIBERO’s four language-conditioned task sets.SimplerEnv probes visual appearance shifts; LIBERO includes Spatial, Object, Goal, and Long tasks on a Franka arm.Source: §4.1 · p. 8

4.1.1. SimplerEnv Benchmark

• InternVLA-M1 improves over Vanilla VLA by 14.6% on Google Robot Visual Matching, 12.4% on Visual Aggregation, and 17.0% on WidowX.Against prior state-of-the-art models, it gains 5.9% on Google Robot Visual Matching, 5.3% on Visual Aggregation, and 9.8% on WidowX.Source: Table 1 · p. 8, Table 2 · p. 9, §4.1.1 · p. 9
• Spatially guided co-training preserves stronger spatial grounding while improving manipulation performance and convergence compared with omitting spatial guidance or vanilla co-training.The spatial-guided approach raises Projection-space Similarity from 0.25 to 0.42; Table 3 compares multimodal understanding, spatial grounding, and manipulation across training strategies.Source: §4.1.1 · pp. 9–10, Figure 5 · p. 10, Table 3 · p. 10

4.1.2. LIBERO Benchmark

• On LIBERO, InternVLA-M1 achieves 98.0% success on Spatial, 92.6% on Long, and 99.0% SR for object placement.The paper reports notable improvements over strong baselines including GR00T N1 and π0, particularly on the spatial and long-horizon tracks.Source: Table 4 · p. 11, §4.1.2 · p. 11

4.2.1. Evaluation in Simulated Large-scale Pick-and-place

• Across 200 simulated pick-and-place tasks, InternVLA-M1 with mid-training gains an average 6.2% over GR00T N1.5 and outperforms π0 in all four settings.Settings are in-distribution, unseen objects, new backgrounds, and unseen instructions; mid-training improves generalization to novel objects, scene shifts, and paraphrased instructions.Source: Figure 7 · p. 12, §4.2.1 · p. 12

4.2.2. Evaluation in Real-world Cluttered-scene Pick-and-Place

• In real-world pick-and-place, co-training with InternData-M1 improves generalization to unseen objects, object positions and orientations, and novel instructions.Both InternVLA-M1 variants outperform GR00T N1.5 and π0 in-distribution; the evaluation includes five regimes and 300 rollouts per model.Source: §4.2.2 · pp. 13–14, Figure 10 · p. 15

4.2.3. Evaluation in Long-horizon and Reasoning Manipulation

• On long-horizon tasks, InternVLA-M1 outperforms baselines and adapts to physical interference and mid-task instruction changes with less performance degradation.Its unified model jointly trains on task decomposition, subtask identification, numerical reasoning, and action supervision; additional post-training improves planning performance.Source: §4.2.3 · pp. 14, 16, 17, Figure 12 · p. 17

Hierarchical robot system

• Prior systems use textual plans, perceptual or 3D cues, scene graphs, and affordances; InternVLA-M1 integrates latent spatial guidance into downstream action training.Examples include textual plans (Zawalski et al. (2024)), detector boxes (Griffin (2023)), grasp points (Ten Pas and Platt (2017)), scene graphs (Rana et al. (2023)), and affordances (Nasiriany et al. (2024)).Source: §5 · p. 18

Embodied reasoning and planning in VLA

• VLA reasoning progresses from linguistic plans and action language to visual or spatial representations, while post-training unlocks intrinsic VLM reasoning without explicit generative steps at inference.Source: §5 · p. 18

Generalist robot policy

• Generalist policies include monolithic VLAs, modular cognition-action systems, and predictive world models; InternVLA-M1 builds on unified VLA with added planning for real-world adaptability.World-model planning can be computationally intensive.Source: §5 · pp. 18–19

6. Discussion and conclusion

• InternVLA-M1 combines large-scale multimodal pre-training with spatially guided post-training to transfer spatially grounded understanding into embodied control.The framework uses spatial grounding priors to bridge high-level multimodal reasoning and low-level robotic execution.Source: §6 · p. 19
• The model generalizes to unseen objects, instructions, and environments, with evaluations spanning simulation and real-world settings.Source: §6 · p. 19
• InternVLA-M1 surpasses existing VLA models and specialized systems in instruction following, long-horizon manipulation, and multimodal grounding.Source: §6 · p. 19