Figure has pretrained Helix 2.5 on Index, its dataset of human behavior, and run three whole-body behaviors — tidying living rooms, folding towels and making beds — in 30 Bay Area homes with no data collected in any of them. The company reports that Index pretraining alone lifts zero-shot success from 9% to 56%, and that Helix 2.5 matches a Helix 02 policy while using half the task-specific data.
Larger pre-trained video models consistently yield better robot policies, and the trend still holds at the largest size tested, the authors report from a scaling study of Direct Video-Action models pre-trained on web video. More pre-training compute improves policies at every amount of robot data tested, with the largest gains when robot data is scarce. The findings rest on over 200 hours of real-robot evaluation on an industrial bearing-unpacking task.
A humanoid from ETH Zürich's Robotic Systems Lab jumps onto a set of monkey bars, traverses them through sparse bar interactions and lands safely. The lab frames the problem as perceiving thin, overhanging geometry while executing agile, accurate whole-body motions. The video leads IEEE Spectrum's Video Friday for the week.
Changing a single environment variable in ROS 2 lets an adversary run a pre-built hook that intercepts and injects both telemetry and control signals before they are published. A robot can then be hijacked into dangerous behaviour while feeding downstream verifiers synthesised telemetry, and the hooks can travel inside third-party Docker containers. On a physical Franka Emika arm running Secure ROS 2, the fabricated telemetry was injected in real time with about 3 ms of jitter, preserving temporal synchronisation.
GE-Act 2.0 trains every parameter from random initialisation on embodied manipulation data, rather than adding action prediction to a pretrained video generator. AgiBot reports that scaling the corpus from 300 to 30,000 hours lifted zero-shot success on its G1-OP evaluation from 17.1% to 44.1%, and the number of tasks with any success from 39 to 76, with towel-folding and pen-recapping appearing only at the larger scales. It says performance had not saturated at 30,000 hours.
ContextFlow learns robot manipulation from demonstrations held in context, using conditional flow matching rather than the next-token prediction that discretises actions and compounds early errors. Perceiver-style compressors distil visual, proprioceptive and action sequences into compact latents, so several demonstrations can be carried at once. On LIBERO it beats ICRT by 35 percentage points on unseen task configurations and matches a fine-tuned π0 without any fine-tuning of its own.
Astribot's SmoothRL runs online reinforcement learning while the robot keeps moving, learning from the actions actually executed in the world. The company reports dynamic tossing rising from 39% to 94%, pen capping from 8% to 83% and box opening from 30% to 90%. Astribot calls it the first reinforcement learning validated on real-world, high-dynamic tossing, and has published uncut footage alongside a technical report.
A researcher shows that a compromised screen displaying a fake system update can make Gemini Robotics 2.0 save a sleeper-agent skill that fires later when it sees a trigger object. In a MuJoCo simulation the robot dog was manipulated into attacking a child, through the same prompt-injection weakness language models have. The attack runs in three steps: the fake update plants the skill, a pineapple in view activates it, and the dog then acts on the hidden instructions.
World Labs has introduced Atlas, a multimodal autoregressive diffusion transformer pretrained from scratch on text, images, video and 3D. All inputs share one spatial context, from which the model generates what comes next while staying consistent in 3D with everything it has already seen. It covers world generation, reconstruction and simulation, including camera-controlled generation from one or more images, and World Labs says performance keeps improving as training compute grows.
North Carolina State University researchers have built a teardrop-shaped liquid crystal elastomer robot that twists under infrared light, stores the energy and snaps itself into a jump, resetting on its own so it keeps leaping while the light stays on. The angle of a stiff V-shaped tail decides whether it crawls, jumps forward or leaps upward; the work is published in PNAS.
CLAP trains action-conditioned video generation across both human and robot footage, reconciling incompatible action spaces through end-effector poses, language and latents. It covers DROID, Bridge, bimanual YAM robots and G1 humanoids, and all code and models are open-sourced. A curriculum first learns physical priors from unlabelled video before specialising, since action representations differ sharply between platforms and are absent from human footage altogether.
SOLO walks a humanoid 1.5 km outdoors zero-shot on nothing but a chest-mounted depth camera and proprioception. It replaces dense terrain reconstruction, which smooths away the detail a footstep needs, with Fourier-encoded cell queries that cut height-map error by a factor of 3.3 to 4. Folding teacher-student disagreement into the PPO reward carries that penalty back to earlier actions, and on stress terrain it reaches 97.5% traversal and 96% on stepping stones against 75% and 0–3% for dense-reconstruction variants.
NVIDIA describes an agent-driven workflow in which a coding agent validates dependencies, prepares assets, runs smoke tests, launches training and compares checkpoints for a navigation policy. Human approval gates control scene acceptance, the smoke test and checkpoint promotion, so the work is not repeated for every robot-scene pair. The tutorial uses Spot as the reference robot across a built-in scene and a SAGE-10K scene, with Omniverse NuRec supporting captured environments.
Generalist AI's GEN-1.5 learns a new task in seconds from a single example, without gradient updates or fine-tuning. The company notes the tasks are simple and short-horizon, but says this is the first model it knows of where one-shot and few-shot learning of physical skills have emerged at scale. It also shows zero-shot physical generalisation, which the company calls a step toward general intelligence for the physical world.
Hydra-0 represents robot actions as pixel motion, giving one visual interface for learning action consequences across embodiments, tasks and video backbones. The best configuration cuts robot-motion error by 90.4% against an action-conditioned baseline, and an emergent inverse mode predicts robot motion from object flow taken from a human demonstration. On the RoboLab benchmark it reaches a Pearson correlation of 0.96 between replayed and reference success rates, and supports zero-shot composition.
Skild AI presents S1 as in-context learning for robotics: one video prompt teaches an unseen task over ten-minute horizons, with no fine-tuning. The company draws the analogy to the shift from BERT-style models, which needed fresh data for each application, to prompting. S1's own framing is that robustly executing a new task has until now meant collecting hours of data and fine-tuning a specialist policy.
Researchers in the Czech Republic built microscopic robots from stacked MXene particles that removed microplastics from both soil and water in laboratory tests. Particles under five millimetres are especially hard to recover from soil because they lodge between minerals and organic matter. The MXene particles carry a large surface area and a surface chemistry that attracts plastic, and a nickel coating lets rotating magnetic fields drive them.
DreamX-Phi 1.0 predicts the frames that follow a prescribed sequence of end-effector poses and gripper states, with per-arm geometric encoding so each arm follows its commanded path, and a depth branch and object masks to keep the scene consistent. It took first place on Track 1 and second on Track 2 of the WorldArena 2.0 Challenge; the model and code are to be released.
World Labs sets out which functional pieces actually compose a world model and what each one is for, arguing the term has become one of the most overloaded in AI. The premise is that language models learn the statistical structure of text while world models must learn the structure of space and time. Computer vision, robotics, reinforcement learning and generative AI each claim to be building world models, and each means something quite different by it.
Flex-π, a 6B-parameter world-action model, is supervised on 3D pointmaps and DINO semantics as well as RGB, all encoded by the same frozen video VAE, and one checkpoint runs on any subset of those streams. The authors report gains of up to 2–7x over the strongest baselines on real bimanual tasks, running faster than π0.5.
A system from MIT uses three agents to assemble the objects, walls and overall appearance of a 3D scene. The generated indoor spaces let robots practise skills and try alternative approaches before they are powered on, addressing how slow and costly it is to teach machines physically. The researchers put the remaining bottleneck in content rather than physics: simulator engines have improved, but generating scenes rich and varied enough to match the real world has not.
NVIDIA argues that a vision-language backbone learns to describe the world rather than predict how it evolves, which is exactly what a robot needs when a task depends on anticipating change. The post sets out world action models as the replacement for that backbone. NVIDIA researcher Jim Fan set out the shift in his Robotics' End Game talk, later condensed to the line that VLAs are dead.
Open-DiffLoco is an open-source framework that trains blind quadruped locomotion policies with differentiable simulation, running the Short-Horizon Actor-Critic algorithm in MuJoCo XLA. Deployed on a Unitree Go2, the proprioceptive policy tracks omnidirectional velocity commands with under 0.2 m/s of error, runs faster than 1 m/s and withstands lateral pushes, without reference trajectories or complex auxiliary rewards. Training takes 20 to 60 minutes on a single RTX 5080 with under 6 GB of video memory.
The authors trace the failures of vision-language-action models in contact-rich manipulation to two causes: precision failures, from a mismatch in how flow-matching policies are trained, and force failures, from the structure of force signals. FACT combines a targeted fix for each and reaches a 66% average success rate on five contact-rich tasks, against 41% for the best prior baseline, over almost 2,500 real-world rollouts.
