Academic

StarVLA-α: Reducing Complexity in Vision-Language-Action Systems

Vision-Language-Action (VLA) models have recently emerged as a promising paradigm for building general-purpose robotic agents, yet the VLA landscape remains highly fragmented in architectures, training data, embodiment configurations, and benchmark-specific engineering. We introduce StarVLA-α, a simple yet strong baseline that deliberately minimizes architectural and pipeline complexity to reduce experimental confounders and enable systematic analysis of design choices including action modeling strategies, robot-specific pretraining, and interface engineering. Across unified multi-benchmark training on LIBERO, SimplerEnv, RoboTwin, and RoboCasa, the same simple baseline remains highly competitive — our single generalist model outperforms π₀.₅ by 20% on the public real-world RoboChallenge benchmark, suggesting a strong VLM backbone with minimal design is already sufficient without additional architectural complexity.

StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing

StarVLA is an open-source codebase that addresses the fragmentation of VLA research across incompatible architectures, codebases, and evaluation protocols. It provides a modular backbone–action-head architecture supporting both VLM backbones (e.g., Qwen-VL) and world-model backbones (e.g., Cosmos) alongside representative action-decoding paradigms (FAST, OFT, flow-matching π, GR00T-style dual-system); reusable training strategies including cross-embodiment learning and multimodal co-training; and a unified evaluation interface across LIBERO, SimplerEnv, RoboTwin 2.0, RoboCasa-GR1, and BEHAVIOR-1K supporting both simulation and real-robot deployment. As a core contributor, I help maintain training infrastructure, benchmark integration, and the community release pipeline.

ST4VLA: Spatially Guided Training for Vision-Language-Action Models

ST4VLA establishes spatially guided training as a unifying principle for scalable, instruction-following generalist robots. Through a two-stage pipeline — large-scale spatial grounding pre-training on 2.3M spatial reasoning samples to determine "where to act", followed by spatially guided action post-training that produces embodiment-aware actions via plug-and-play spatial prompting — the recipe yields consistent gains across SimplerEnv Google Robot (+14.6%), WidowX (+17%), and LIBERO Franka (+4.3%) while strengthening box, point, and trace prediction. With 244K simulation pick-and-place episodes for co-training, it achieves +20.6% on unseen objects and surpasses prior work by over 10% on long-horizon reasoning-intensive tasks.

SynthVerse: A Large-Scale Diverse Synthetic Dataset for Point Tracking

SynthVerse is a large-scale, diverse synthetic dataset specifically designed for point tracking. It introduces several new domains and object types missing from prior synthetic datasets — animated-film-style content, embodied manipulation, scene navigation, and articulated objects — substantially expanding diversity while providing high-quality dynamic motions and interactions. We further establish a highly diverse point tracking benchmark for evaluating state-of-the-art methods under broader domain shifts. Extensive experiments show that training with SynthVerse yields consistent improvements in generalization and exposes limitations of existing trackers under diverse settings.

Nimbus: A Unified Embodied Synthetic Data Generation Framework

Nimbus is a unified synthetic data generation framework integrating heterogeneous navigation and manipulation pipelines for embodied intelligence. It introduces a modular four-layer architecture with a decoupled execution model that separates trajectory planning, rendering, and storage into asynchronous stages. Through dynamic pipeline scheduling, global load balancing, distributed fault tolerance, and backend-specific rendering optimizations, Nimbus achieves a 2–3× improvement in end-to-end throughput over unoptimized baselines and serves as the production backbone for the InternData suite.

InternData-A1: Pioneering High-Fidelity Synthetic Data for Pre-training Generalist Policy

InternData-A1 is a large-scale synthetic robotic dataset (630k trajectories, 7,433 hours across 4 embodiments and 70 tasks) generated through a fully autonomous and compositional simulation pipeline. Using the same architecture as π₀, we show—for the first time—that a VLA model trained entirely on synthetic data can match the strongest real-robot datasets, achieving comparable performance across 49 simulation tasks, 5 real-world tasks, and long-horizon dexterous manipulation. The model also demonstrates zero-shot sim-to-real transfer, highlighting the substantial value of scalable simulation for embodied AI. The dataset and generation pipeline are released to enable broader access to large-scale robotic data creation.

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

InternVLA-M1 is a unified framework for spatial grounding and robot control that advances instruction-following robots toward general-purpose intelligence. Its core idea is spatially guided vision-language-action training, where spatial grounding serves as the critical link between instructions and robot actions. InternVLA-M1 employs a two-stage pipeline: (i) spatial grounding pre-training on over 2.3M spatial reasoning data to determine “where to act” by aligning instructions with visual, embodiment-agnostic positions, and (ii) spatially guided action post-training to decide “how to act” by generating embodiment-aware actions through plug-and-play spatial prompting.

GenManip: LLM-driven Simulation for Generalizable Instruction-Following Manipulation

GenManip is an LLM-driven, Isaac Sim–based embodied manipulation benchmark with automatic demonstration and layout generation and closed-loop evaluation. Designed to systematically test generalizable instruction-following manipulation in the era of MLLMs, it has served as core infrastructure supporting subsequent VLA research at Shanghai AI Laboratory.

PMT: Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation

Proposed a semi-supervised learning framework for medical image segmentation using Mean Teachers to enhance model diversity and regularization. Achieved state-of-the-art results and demonstrated generalization across datasets.