DragMesh, proposed by a Peking University team, has been accepted by the International Journal of Computer Vision (IJCV). With a single drag on an existing 3D asset, it can recover the joint type, axis, and origin of the target part and generate a complete articulated motion trajectory.
The paper proposes ToM2Vec, which forms abstract theories of mind (ToM) in a latent space through a fixed budget of continuous reasoning tokens. During training, the partner's actual thoughts serve as supervision; at test time, these tokens condition the agent's thinking and actions.
iFlytek has released the end-cloud integrated large model iFlytek Spark X2.6, comprising the cloud-side 293-A30B model and the already open-sourced device-side X2.6-4B model. The cloud model supports a one-million-token context, with an input price of 0.9 yuan per million tokens and an output price of 3.6 yuan per million tokens, lower than DeepSeek-V4.1-Flash-1017.
NEURA Robotics and the Technical University of Munich (TUM) have opened TUM RoboGym at LabCampus near Munich Airport, covering more than 2,300 square meters. TUM MIRMI is a co-builder, the initial collaboration runs for three years, and the two parties are jointly investing about 17 million euros, of which NEURA is contributing about 11 million euros.
Why it matters: The original text covers both the new robot's hardware changes and the three sources of training data, letting readers understand NEURA's approach of tying data infrastructure to the hardware launch.
Dream Labs announced this week that it has raised a $90 million seed round and launched the Dream API, which converts robot camera feeds and natural-language instructions directly into motor commands. Its flagship model, Dream-1, is trained on videos of everyday human tasks; the company says it can control new gripper-equipped robots zero-shot without additional data. It currently supports tasks such as tabletop pick-and-place and clothes folding, with wheeled mobile robots and dexterous hands to be supported later.
TaskAnchor connects execution history to VLA's existing language and vision interfaces through a lightweight adapter, aiming to alleviate task-state confusion in long-horizon manipulation. It uses a dual-path design: a normalized task-state coordinate that describes semantic execution stages, and a visual residual based on context-adaptive test-time training (CA-TTT) that preserves historical details.
The China Telecom Artificial Intelligence Research Institute (TeleAI), Gamma Intelligence, Shanghai Jiao Tong University, the University of Hong Kong, and Zhejiang University jointly proposed NEXUS, a cross-terrain perception and motion control foundation for humanoid robots that lets robots autonomously adjust posture and contact based on terrain while preserving motion intent.
A joint team including SenseTime and The Hong Kong University of Science and Technology (Guangzhou) released the report "Has GPT-6-Astra Achieved Embodied Intelligence?", systematically evaluating GPT-6-Astra's embodied capabilities across 12 benchmarks, and simultaneously open-sourced the SenseNova-RoboRSI evaluation framework and a 27B embodied model.
Why it matters: The report breaks model capabilities into testable conditions and provides harness ablation data, so readers can judge the real bottlenecks in embodied capability evaluation and cost.
Embodied intelligence evaluation platform RoboColiseum released an evaluation report: Claude Opus 5.5 ranked first among four multimodal models with 76.28 points, followed closely by GPT-6 Astra with 75.99 points, while GPT-6.1 Sol and Gemini 4 Argon scored 70.91 and 66.86 points respectively; the top two on the official overall leaderboard remain specialized embodied policy models Paimon (77.54 points) and HiDream-O1-Embodied (76.92 points).
Tesla reported Q3 2026 revenue of $28.046 billion and net profit of $787 million, down 43% year over year. On the earnings call, Musk said Optimus has entered production on a line at the Fremont factory and is expected to be delivered to external customers by the end of next year, reaching an output rate of 20,000 units per week by the end of 2027. He also said an "Optimus Academy" will let the robots observe human tasks and practice repeatedly in simulation and reality to form a reinforcement-learning flywheel, and that the AI 5 chip will be integrated into Optimus around April next year.
At the 2026 Chery Global Innovation Conference, Chery released CheryMind, a vertical foundation model for physical AI. It serves as the technology foundation of the Yaoguang 2035 strategy, covering automotive R&D and manufacturing and extending to businesses including intelligent driving, intelligent cabins, robotics, eVTOL, and agricultural machinery.
The study proposes SERA (Self-Evolving Robotic Agent), which organizes a high-level VLM into a robotic agent through an editable harness, with a fixed VLA executor providing low-level control. The harness uses observations and execution feedback to maintain task context, and the self-evolution mechanism proposes and evaluates harness revisions through environment rollouts, with the selected configuration fixed before held-out testing. On RoboMME, SERA autonomously selects a configuration reaching 51.38% overall success rate, versus 44.38% without self-evolution.
A team led by Dawei Zhou at Virginia Tech and researchers at Meta propose ReGDiff in a NeurIPS 2026 paper to generate metamaterial voxel structures that satisfy geometric constraints. The method has two steps: first, an RAS mechanism organizes the autoencoder latent space into a form where good and bad structures are clearly separated and good structures are evenly distributed; then, short-range repulsion (SRR) pushes latent variables away from the near neighbors of known samples during sampling, and it takes effect only at inference time.
NVIDIA announced several implementations of open models and physical intelligence in European retail: Shopify used NVIDIA Megatron Bridge and NeMo RL to post-train Nemotron 3.5 Lightning, cutting serving costs by 90% and delivering 7x faster responses on internal benchmarks.
The paper finds that VLA action generation differs from text-to-image generation: merely biasing the training timestep distribution toward the noise endpoint yields strong one-step policies on LIBERO, LIBERO-Plus, LIBERO-Pro, and real-world robot tasks. For open-source VLA models with different architectures, directly replacing iterative decoding with a single forward pass often performs comparably to the original multi-step decoding on LIBERO, CALVIN, and SimplerEnv.
PaceVLA proposes an adaptive temporal execution framework for VLAs that uses multi-granularity action decoders on a shared action expert, along with quantizability-guided mode selection and context-conditioned routing, to decide at each inference step whether to adopt coarse-grained execution. Experiments cover three multi-task simulation benchmarks and six real-world tasks, improving execution efficiency while maintaining task success rates across multiple embodiments, including long-horizon and dexterous manipulation.
The paper proposes the Collective Experience Learning (CEL) paradigm, in which deployed robots pool their own execution experience to update a shared policy that is then returned to the fleet for use. It is implemented through failure-guided experience retrieval fine-tuning (FERT), which uses failure contexts to retrieve relevant segments from successful trajectories and prioritizes their recorded actions for supervised updates, without additional expert demonstrations or corrective actions; a supporting fault-tolerant asynchronous training system (FATS) coordinates robot execution, cloud inference, and policy optimization.
Why it matters: The paper proposes replacing expert corrections with the deployed robots' own execution experience to iterate on the policy, and reports real-world manipulation results at a scale of 96 robots.
The paper proposes Vector-Quantized World Model (VQ-WM), a latent feedback motion planning framework that combines discrete long-horizon search with continuous visual feedback control: a discrete latent state predictor models transitions between vector-quantized latent states, allowing the planner to search for the shortest path over long distances, while a continuous learned controller then reaches the planned states step by step through visual feedback to correct deviations and mitigate compounding errors.
FoveaVLA is a vision-language-action (VLA) model that replaces the wrist camera with learned foveation actions. Alongside each action chunk, it predicts the interaction region to observe at the next policy query, then crops that region from the original-resolution image and feeds it to the policy together with the full view.
The Lambda-humanoid Team released λ0, a humanoid whole-body loco-manipulation foundation model, along with HumanVerse, a 500-hour open-world human loco-manipulation dataset that pairs first-person video with synchronized body and hand motion.
HarnessPAI proposes a model- and embodiment-agnostic Harness framework for Physical AI that treats code as the executable, evolvable interface organizing underlying action primitives. The framework separates two timescales: within a single rollout, it executes open-loop at the program level with a fixed program guiding and checking execution; across rollouts, it evolves closed-loop, using execution feedback to revise the program and distill failures into reusable skills.
The paper proposes BAS-VLA, a task-semantic action calibration framework built on top of a frozen base VLA, designed to address two types of failure: policies producing unnecessary action drift when task semantics are unchanged but appearance varies, and policies continuing along the original trajectory when key semantics such as the target object or constraints are altered.
The paper proposes rMuscle, a real-time vision-language-action (VLA) model inference framework inspired by human muscle memory. The authors first analyze embodied workloads and find that when robots repeatedly perform tasks, there is significant similarity in observations, action trajectories, and even internal model states.
This study proposes a local action-conditioned world model framework for probe guidance toward target planes in robotic ultrasound, using a two-stage model-based learning pipeline: first, a latent conditional diffusion world model predicts future ultrasound observations from recent context frames, probe motion, and temporal offset; second, a goal-conditioned temporal Transformer predicts ordered probe actions and is fine-tuned using rewards from the frozen world model.
ReStruct proposes a method to steer learned robot policies at inference time, enabling them to satisfy user preferences not specified during training without retraining. The method decomposes a visuomotor policy into a high-level state-machine skeleton and a low-level residual controller, and incorporates preferences into the skeleton via a synchronous product to reconfigure the task structure, while the controller remains frozen. Simulation and real-world experiments show that ReStruct can handle a variety of preferences, from object-level specifications to temporal-logic constraints, and after steering it outperforms existing methods in task success rate and preference adherence, leading VLA models by up to 25%.
SafeDojo proposes a model-based safe reinforcement learning framework that improves vision-language-action (VLA) policies through world-model imagination, avoiding costly real-world exploration or reliance on hand-crafted safety functions. The method performs online reinforcement learning on top of an interactive video world model, using a ResNet success classifier to score predicted future frames, paired with a lightweight safety head that predicts per-step contact probabilities from the current latent context and candidate action chunks, then balances task execution and trajectory safety signals through Lagrangian-constrained GRPO.
μ0 is a scalable world model based on 3D traces. Rather than predicting dense pixels or directly modeling actions, it forecasts smooth 3D trajectories for salient interaction points such as objects, tools, hands, and contact regions, forming a compact, embodiment-agnostic motion interface.
MiTaS is a multi-resolution tactile representation framework that uses modality-specific convolutional stems and a Transformer-based fusion structure to integrate information from an RGB camera, a GelSight Mini vision-based tactile sensor, and a high-frequency event-based Evetac sensor, then conditions a flow-matching policy to perform downstream tasks.
This paper studies the dynamic robot allocation problem in heterogeneous multi-team collaboration, treating robots as transferable resources, drawing on Hamilton's rule from ecology to design an altruistic decision-making mechanism, and proving that the generalized heterogeneous resource allocation problem is combinatorial optimization and NP-hard.
This paper proposes a feedback world model that uses the true states observed after robot execution to perform online closed-loop correction of predictions at inference time, without additional training data or parameter updates. The method can be interpreted as a latent-space observer and provides convergence guarantees under bounded, asymptotically convergent dynamics residuals and a stable feedback gain; it also introduces action-aware guidance to better translate corrected predictions into control.
Move-Then-Operate is a vision-language-action (VLA) framework that explicitly decouples robotic manipulation into two behavioral phases: coarse relocation (move) and contact-critical interaction (operate), using a dual-expert policy routed by a learnable phase selector.
Safe-Night VLA is a multimodal manipulation framework that integrates long-wave infrared thermal perception into a pre-trained vision-language backbone, enabling robots to perform semantic reasoning based on thermodynamic properties and to enforce deterministic workspace constraints during policy execution through a control barrier functions safety filter.
AgiBot World Colosseo introduces the AgiBot World large-scale manipulation platform, containing over 1 million trajectories covering 217 tasks across 5 deployment scenarios, an order-of-magnitude increase in data scale over existing datasets, and uses a standardized collection pipeline with human-in-the-loop verification.
This study examines the degradation of world-model predictions when physical conditions fall outside the training distribution, comparing two causal modeling elements: the inputs available to the predictor and how its mechanism computes their effects. On a synthetic benchmark with shortcut features, a frozen dependency mask derived from paired interventions reduced the out-of-distribution RMSE of a fully connected predictor from 1.769 to 0.697, close to the generator-oracle mask's 0.694.
The paper proposes AW-0, an agentic world action model framework that combines code-based planning with joint video and action prediction. Within a unified mixture-of-transformers architecture, it simultaneously models three coupled generative streams: code, future video, and robot actions. A code expert generates parameterized skill routines and control logic, while video and action experts generate future visual states and low-level motor commands.
PhysiCalWorld proposes calibrating action-conditioned world models' physics via in-context prompting, without fine-tuning or analytical physics modeling, to adapt to instance-level dynamics.
CW-WAM is a world action model that jointly predicts video, motion, and contact wrench. Inspired by hybrid motion-force control, it uses a unified interaction module to predict time-aligned motion-wrench chunks as control references. It structures motion-wrench dependencies with pair-causal attention and retains past interaction information in a fixed-length wrench memory; the predicted wrench chunk serves as the force reference for a 20 Hz outer-loop admittance controller.
The paper proposes PAPAV, a capability-centric unified framework that distills the recurring functions of multimodal embodied agents (MMEAs) into five: Perceive the current state, Anticipate action effects, Plan a feasible course, Act through an interface or body, and Verify the outcome.
The paper proposes AtomicRL, a framework that uses atomic task structure to improve reinforcement learning post-training of vision-language-action (VLA) policies. Its core components include Semantic-Aware Policy Optimization (SAPO), which uses semantic gating to prevent TD residuals from failed atomic subtasks from propagating across tasks, thereby mitigating credit contamination; and Bottleneck-Adaptive Target Exploration (BATE), which dynamically coordinates task sampling and local exploration incentives based on the mastery level of each subtask.
The study trained 12 policies with the same recipe on a simulated Unitree G1, varying gravity, terrain geometry, and ground contact models (rigid or 3D resistive force theory, 3D RFT), and tested Earth-to-Moon fine-tuning and gravity randomization.