Robotics & Physical AI

TacForcing is a streaming action-generation framework that lets vision-language-action policies use tactile feedback as actions are being executed. It replaces a separate high-frequency reactive controller with a streaming action expert and reports stronger average success on simulated and real contact-rich tasks.

Zero-WAM treats a human manipulation video as an in-context task specification for a robot, rather than requiring parameter updates for every new task. The project introduces the HumanGen dataset and reports a 47.0% average success rate on seven unseen RoboTwin 2.0 tasks.

Anthropic has introduced a research preview of its Model Hardware Standard, or MHS, a proposed common interface for AI agents to discover and operate robots, microscopes, cameras, and laboratory instruments. Rather than giving Claude one permanent body, the approach lets it use many connected devices as temporary extensions.

StreamPI adds temporal reasoning to single-frame vision-language-action models without introducing extra parameters. Its instruction-anchored attention and randomized interval training are designed to improve memory, spatial perception, and robustness to asynchronous robot inputs.

GigaBrain-0.7 proposes a three-system architecture that connects visual-language understanding, world prediction, and robot action. Trained on more than 37,000 hours of heterogeneous embodied data, it targets stronger transfer across tasks and robot embodiments.

WorldToken combines multiview images, proprioception, and task conditioning into one token for each policy timestep, then models the resulting sequence causally. Its experiments point to the importance of target-domain data and long temporal context in robotic imitation learning.

PhysCaP extends code-as-policy robotics with physics-informed active perception. Instead of relying only on what a camera can see, the agent uses targeted interactions and proprioceptive feedback to infer hidden properties such as mass and stiffness.

General Intuition is reportedly discussing a new financing round at a $6 billion pre-money valuation. The startup plans to use the capital to advance its action-focused foundation model and robotic embodiments.

τ₀-VLA introduces world-model-guided test-time computation for long-horizon robot manipulation, allowing a high-level policy to search through candidate subtasks when a decision is uncertain.

DeepMind introduces EXIMO, a three-stage method that combines VLM-guided exploration, imitation learning, and residual off-policy reinforcement learning to fine-tune vision-language-action policies for new tasks.

Jiushi Intelligent and Yutong have launched the Z20, an autonomous light truck rated to carry 4.2 tons with a 19.32-cubic-meter cargo box. Its first targets are factories, industrial parks and large warehouse hubs rather than complex urban roads.

Generalist’s GEN-1.5 brings context learning into the physical world. The model can attempt some unseen tasks after watching a 3–12-second demonstration, without updating its parameters.

Zetta is a closed-loop embodied harness that evolves runtime critics and recovery skills while leaving the base policy frozen. Its three-level design connects action-time intervention, rollout analysis, and validation-gated updates.

Ego2Robot is a data synthesis pipeline that transforms egocentric human manipulation videos into robot-style training data. The paper argues that this kind of synthetic data can support large-scale pretraining, not just narrow task retargeting.

The paper argues that many autonomous-driving VLM training pipelines let teacher models see the ground-truth future trajectory too early, encouraging post-hoc rationalization rather than causal reasoning. AD-MCQ and DEFT-RLVR recast planning as verifiable trajectory selection.

N₀-TWAM is a tactile-native world-action model for contact-rich manipulation, designed to predict not only future vision but also future contact before generating actions.

SGTP combines game-theoretic reasoning with GPU-accelerated sampling to plan competitive yet feasible behaviors for autonomous racing with multiple interacting vehicles.

A new paper introduces Collective-State JEPA, a decentralized predictive architecture that lets each robot in a swarm form a representation of the same future collective state. The results highlight label efficiency, topology transfer, and planning-relevant value estimation.

RL²-VLA introduces an inference-time steering framework for VLA models that activates reinforcement-learning guidance only when the base policy is likely to fail.

N_0-VTLA targets a persistent weakness of vision-language-action robotics: fine-grained manipulation under rich physical contact. It adds scalable tactile pretraining and an offline improvement method called ALTER.

WAIC 2026 was crowded with robots, foundation models and AI applications, but the more important question was whether these technologies are changing real workflows or merely producing better demos.