World Models

TEMPO Gives a VLA a Short Memory for Dynamic Manipulation

A VLA looking at a single frame cannot tell which way a moving bottle is heading. TEMPO gives the policy a short memory: a frozen video foundation model summarizes recent scene motion and a compact proprioceptive history disambiguates visually similar moments, lifting dynamic bottle handover from 44% to 74% success.

TEMPO: Adding Temporal Context to a Pretrained VLA for Dynamic Manipulation

A VLA sees a single frame, so it cannot anticipate where a moving object is heading. TEMPO adds two temporal inputs to a pretrained VLA without touching the backbone: a motion summary from a frozen video foundation model, plus a compact proprioceptive history. On a bimanual workstation with two 7-DoF I2RT YAM Ultra arms and three cameras, Bottle Handover success rises from 44% to 74% and Flick Catch reaches 66%.

HIL-UMI: Bringing Human-in-the-Loop Post-Training of Vision-Language-Action Models to Universal Manipulation Interface

Large-scale vision-language-action (VLA) models provide powerful priors for robot manipulation, yet adapting them to a specific deployment remains challenging. Supervised fine-tuning (SFT) on task-specific demonstrations provides a step toward deployment, but faces two persistent limitations: static data provide limited coverage of out-of-distribution states, and standard imitation objectives do not distinguish progressing behavior from less useful data. Interactive post-training can address these limitations, but typically requires repeated policy execution and human intervention on a physical robot. We introduce HIL-UMI, a policy-guided Universal Manipulation Interface (UMI) framework for robot-free human-in-the-loop VLA post-training. During handheld UMI demonstrations, HIL-UMI queries the current policy on the same observation stream without executing its predictions. The Energy Score compares the human action trajectory with policy inference and triggers collection when their discrepancy indicates an out-of-distribution region. In a separate feedback loop, low online advantage predictions identify essential segments for refining a progress-based advantage estimator. The updated estimator then guides advantage-conditioned behavioral cloning using a balanced mixture of base demonstrations and new policy data. This design preserves the iterative and policy-aware nature of human-in-the-loop learning while decoupling data collection from robot deployment. Experiments on four real-world tasks spanning long-horizon and precise manipulation show that HIL-UMI achieves consistent improvement over SFT and benefits from both targeted collection and advantage refinement. Moreover, HIL-UMI outperforms HG-DAgger on Clean Up Table with lower per-frame collection time, suggesting a scalable path for VLA post-training across operators and locations.

Han, Zimu, Zeng, Yiming, Zhang, Jiyao

JEPA-Anything: Learning Predictive Models across Different Worlds

JEPA-Anything introduces orthogonal predictive factorization for domain-agnostic world modeling, splitting latent targets into complementary factors with dedicated predictive pathways. Across seven domains, it improves matched dynamics tasks and supports intervention, OOD generalization, and long-horizon forecasting.

Cui, Taoyong, Wang, Zhongyao, Xu, Xinyue

X-WBC: A Cross-Embodiment Foundation Model for Humanoid Whole-Body Control

X-WBC changes the training unit from one robot to one motion: human motion, 5-point VR tracking and robot-specific reference motion all map into a shared motion-token space, a causal Transformer learns the temporal structure, and lightweight per-robot encoders/decoders handle each body. Joint training across nine embodiments gives 94.2% same-motion Recall@1, and the same sparse VR command runs on G1, R1, H1-2 and H2.

XPeng IRON and the XPACE World-Action Model

XPeng unveiled XPACE, a world-action model that learns actions and future video together, alongside plans to mass-produce IRON by year-end.

Odyssey-3: World Models as Training Grounds for Agents

Odyssey-3 renders interactive generated worlds driven in real time by keyboard and mouse: a rabbit in a meadow, first-person cycling, a warehouse picker among shelves. If a world model can simulate infinite, intelligent worlds, it becomes the place where agents themselves are trained.

Odyssey-3: A Foundation World Model for Physical Agents

Odyssey-3 is an autoregressive diffusion-transformer foundation world model that controls robot arms, humanoids, cars, and drones with task-specific experiential data.

CHORUS: One VLA Policy for Robot Collaboration

Stanford CHORUS turns a pretrained VLA into a shared decentralized policy for heterogeneous robot teams without central control or inference-time communication.

Dream-RSI: Recursive Self-Improvement through Evolving Worlds

Recursive self-improvement is becoming increasingly vital for autonomous AI agents, where progress hinges on discovering high-value solutions across complex domains. The driver of this process is effective exploration, however, managing and improving exploration strategies remains a major bottleneck. Current systems face a fundamental dilemma: fixed strategies fail to adapt as search spaces scale, while online policy optimization requires navigating vast meta-search spaces under delayed and expensive feedback over long-horizon rollouts. We introduce {Dream-RSI}, a framework for scalable and recursively self-improving exploration. A lightweight orchestration layer makes exploration explicit and programmable while leaving the underlying coding agent unchanged. Our key insight is that accumulated discovery history can serve as a replay simulator over the realized search space. By performing dreaming in the replay simulator constructed from historical discovery trees, {Dream-RSI} secures immediate, low-cost off-policy feedback to evaluate and refine exploration policies without invoking repetitive, expensive online evaluations. The improved policy is subsequently redeployed online to drive further discovery, continuously expanding the simulator pool in a self-improving loop. Across algorithm engineering, mathematical optimization, and GPU kernel engineering, {Dream-RSI} achieves competitive or improved discovery quality while substantially reducing discovery cost in several settings.

Tong Zheng, Xidong Wu, Zheng Zhang

reBot Leader Arm Mirrors Teleoperation into Isaac Sim

Seeed Studio's reBot leader arm mirrors a LeRobot teleoperation session into NVIDIA Isaac Sim over UDP, supporting data collection and VLA training.

FTP-1: A Generalist Foundation Tactile Policy

FTP-1 unifies image-, array-, and state-based tactile sensing through a Morphology-Aware Tactile Token Space, pretrained on about 3,000 hours of heterogeneous manipulation data.