NVIDIA just dropped a crazy new research idea. Get ready. The same next-token trick behind ChatGPT got pointed at a ragdoll. Honestly the result is unsettling. No exotic new architecture. Chop human motion into a vocabulary of tiny movement tokens. Train a plain GPT to guess the next one, decode it into muscle commands, let physics run. Then the strange part shows up. Shove the character and it catches its own step. Knock it flat and it plants a hand and gets back up. NOBODY wrote a reward for standing up. Here is why this rattles robotics people. Every humanoid team right now is buried under hand-tuned controllers, one for each little skill. 🔵 the old way: script every skill and its reward by hand 🔵 this way: pretrain ONE body on a mountain of motion and the instincts come free Falling and getting up was never programmed. It emerged. Movement is just a language you can autocomplete.

Agility Robotics is going public. Company known from their humanoid robot Digit and Toyota factory deployment. It is a SPAC merger with Churchill Capital Corp XI, run by Michael Klein. The deal values Agility around $2.5B and raises over $620M in gross proceeds. That is the largest capital raise ever tied to a humanoid company going public. It also makes Agility the first pure-play humanoid company to trade on public markets. 🔵 Figure AI raised $1B at a $39B valuation and stayed private. 🔴 Agility is going public at $2.5B The debut is backed by REAL REVENUE. Over $300M booked, multi-year, tied to roughly 1,000 robots on a robots-as-a-service model. Clients already include GXO Logistics, Amazon, Toyota, Schaeffler, Mercado Libre. Agility has also cleared industrial safety certification to run inside customer facilities. The deal still needs shareholder approval and SEC review, expected to close later this year.

UBTECH’s Walker S2 deplyoed into a busy China–Vietnam border hub at Fangchenggang. The contract behind it is $37M This is a high-friction operations environment: 🔵 passenger queues 🔵 customs questions 🔵 cargo verification 🔵 corridor patrols 🔵 crowd density monitoring 🔵 data handoff to human reviewers Chinese reports describe the robots as assistants, not final decision makers. Walker S2 was chosen because of its autonomous battery-swapping system for continuous operation. Target use cases include airports, ports, rail stations, logistics hubs, and large public venues. The system is designed for integration into public-sector workflows and data collection processes.

China is doing to robots what it did to EVs and drones. From January to May, China exported 10.3M robots with nearly 20B yuan (~$3B) in export value. Most of that is not humanoids. 🔴 cleaning robots 🔴 industrial robots 🔴 cobots 🔴 inspection robots Here’s what actually drives those numbers: 🔵 cleaning robots dominate over 70% of export value 🔵 around 70,000 industrial robots were exported in 5 months 🔵 China became a net exporter of industrial robots in 2025 🔵 intelligent bionic robot exports passed 8,000 units The 2025 report shows that China has 140 complete robot manufacturers.

Make your own humanoid robot! Here is top 5 fully open sourced humanoid robots: 🔵 Asimov 1 — $15K target DIY kit. Full open-source stack. 🔵 Berkeley Humanoid Lite — sub-$5K. Open-source 3D-printed humanoid. 🔵HopeJR — ~$3K. Fully open-source humanoid from Hugging Face 🔵 ROBOTIS K0 — price not announced. Full size open-source humanoid platform. 🔵 Vibe A1 — Mini humanoid. Sales price starting at $649.

Roblox engineer Tom Sanocki joins 1X Gaming talent flowing to robotics to build world models!

Lightwheel builds robotics sim & data infra Raised $145m, based in China 🇨🇳 🔴 Turns physical objects into sim assets 🔴 Collect high-quality egocentric data 🔴 Evaluate data in sim environments They partner with top AI labs (OpenAI, NVIDIA, Alibaba, ByteDance, Manus, Deepmind) and manufacturers (Figure, Toyota, GM, Samsung, Galbot, Agibot) I talked to @jonstephens85 from @LightwheelAI

This team built the robot around the data problem, not the other way around. Human-as-Humanoid, new work from HKUST(GZ) and DeepCybo. The goal: turn plain human videos into REAL training data for humanoids. 🔵 Most efforts collect videos first, then fight the human-robot gap in software. 🔵 Here the robot itself is built to match human proportions. Same reach, same hand size. When the robot already fits inside human motion, the gap almost disappears. The ego exo trick is the clever part. A head camera records what the robot will actually see during deployment. Outside cameras watch the hands without occlusion and recover the motion. One camera cannot do both jobs. Two views, two roles, problem solved. The result: human video becomes robot actions in near real time. Roughly 5x FASTER than collecting the same data through teleoperation. And the robot performed new tasks with ZERO robot demonstrations. Human video only.

You can train a robot policy inside a world model. It just slowly drifts away from reality while you do. WorldSample takes the opposite bet. The RL stays on the REAL robot. The world model is NEVER the environment. It expands one real rollout into many instead. 🔵 It perturbs the actions the robot actually ran, then predicts what happens next. 🔵 A reward model tags each synthetic trajectory success or fail. 🔵 One physical rollout becomes a whole batch of grounded synthetic ones. A world model finetuned only on demos barely moved. Average success went 56% to 82%, with 59% fewer training steps. Base video model is Cosmos-Predict2.5.

Two minutes of data. No rewards, no mocap, no demos. Zero-shot failures turn into 90% success. New from CMU's LeCAR lab: FADA, few-shot adaptation for humanoid control. When a robot fails on a slope or under a payload, its plan is usually still correct. What breaks is execution. The same command no longer produces the same motion under new physics. So FADA splits the policy in two. A planner predicts what the body should do next. An inverse dynamics model turns that into actions. At deployment they freeze the planner and finetune only the IDM with LoRA. Supervision is just 2 minutes of the robot's own rollouts. The paired observations and actions. The rollouts don't even need to be good. Failed attempts work as training data. Results on real hardware: G1 slope traversal goes 0% to 80% success. Booster T1 pulls a 6 kg laundry basket across the line, 100%.

This robot has no touch sensors. It imagines what things feel like instead. New work from Purdue and Texas A&M, called TacImag. Tactile sensors like GelSight are great but fragile. They need calibration, wiring, and they wear out. So the team trained a diffusion model on paired vision and touch data. At deployment the sensor comes off. A dummy finger goes on. The robot watches its own hands and generates the touch signal from vision alone. On real hardware the numbers are wild. Bulb installation jumps from 26.7% to 86.7% success. One intresting thing: the imagined tactile image actually shows golf ball dimples when the gripper grabs one. This isn't magic. It only works when the camera can see subtle contact cues. The model doesn't invent touch. It translates vision the policy couldn't read on its own.

Check this out. This full-size humanoid can carry up to 24 kg without breaking a sweat. Everyone working on humanoids knows the annoying part. Teleoperation looks clean when the robot is just copying motion. It gets much harder when the robot is actually carrying weight. A payload shifts the robot’s balance and makes locomotion less forgiving. Even a small VR tracking error can turn into a real walking problem. That is where HEFT comes in. 🔵 175 cm, 65 kg full-size L7 humanoid 🔵 5 kg loaded basket pickup and carry 🔵 10 kg asymmetric water bottle carry 🔵 two 12 kg kettlebells while walking 🔵 24 kg total payload while squatting It trains on raw VR commands, but uses cleaner reconstructed motion as guidance. The deployed policy still runs from normal VR input. A load that is fine while standing can break the motion during a turn or squat. So each motion is split into short windows. Each window gets its own feasible payload range during training. It is proof that humanoid robots can operate in heavy-duty environments through effective teleoperation

Okay this is kind of wild. You can build full 4D scene reconstruction with zero training. And it works from one ordinary camera. No depth sensor, no multi-view rig. One Video, One World's trick is a clever vision pipeline, not some giant trained model. → A vision-language model scans the video and finds every object. → It tags each one as still, moving, or actually deforming. → Every object becomes a clean 3D mesh, hidden parts filled in first. → It recovers real metric scale and tracks each object frame by frame. → Everything drops onto an estimated ground plane, nothing floats or clips. The output is watertight, separated meshes, ready for a physics engine. Not a render you can only spin around and look at. No rigging, no category priors. Motion is just vertices moving. On video it runs one to two orders of magnitude faster than the alternatives too.

Alibaba just joined the World Action Model race. ABot-M0.5 is their new WAM for mobile manipulation. It is built for robots that need to move and manipulate in the same task. So instead of only controlling a robot arm on a table, the model also handles the robot base. →Go to the object. →Change the camera view. →Predict what comes next. →Then use the arm. ABot-M0.5 does this with a simple pipeline: video>latent action>robot action The other difference is action separation. Base movement and arm manipulation are not treated as one mixed action space. They are separated inside the model, but still trained together. 🔵 future video prediction 🔵 latent actions between video and control 🔵 separate branches for base movement and arm control The thing they call Dream Forcing is intresting. During deployment, the robot will not see perfect future frames. It will act from its own predicted future. So ABot-M0.5 trains the action model on those predicted frames too.Not only on clean ground-truth video.

A robot can now assemble IKEA furniture. Actual table, shelf, and chair assembly with two robot arms. FurnitureVLA frames this as a long-horizon VLA problem. Instead of pushing the robot through one giant rollout, it cuts the build into language-grounded stages. 🔵 the robot follows one clear instruction at a time 🔵 it predicts both the next action and stage progress 🔵 high progress triggers retreat and the next instruction Retreat is doing real work here. They do not switch stages while the part is still in contact. Insertion is noisy, contact is unstable, and tiny alignment errors can poison the next step. After the arms move away, the next stage starts from a cleaner state. Small caveat: no screws yet. They use magnets after alignment, so the Allen key is still alive. Still a very clean step toward home robots handling long, physical tasks without falling apart halfway through.

DexMate raised $33m to build wheeled humanoids I talked to the team at Automate last week 🔵 The large robot is for industrial use cases, has 20hr battery life and experiences minimal overheating 🔵 Their robots are used by top US AI firms and universities for R&D 🔴 HQ in Santa Clara 🇺🇸, manufacturing in Asia 🇨🇳 🔴 Investors include LG and RoboStrategy 🔴 4th largest @RoboStrategy portco @DexmateAI

Weave Robotics launched Isaac 1 for home chores. Deliveries begin this fall. $7,999 upfront or $449/month. 5 different colors. Daily reset feels like the next step after robot vacuums. Instead of just cleaning the floor, Isaac 1 tries to understand what a room should look like after humans ruin it. 🔵 dirty clothes should be gone 🔵 toys should go back somewhere 🔵 beds should look made 🔵 pillows should not be on the floor 🔵 shoes should not live in the middle of the room Their previous robots already folded 1,000 lbs of laundry per week. Isaac 1 is a significant step up from the older model in terms of home chore coverage. Teleoperation support is still included, but at this stage it may be necessary for near -100% task completion.

A robot can now assemble a motherboard. Not in a clean pick-and-place toy setup. It places the CPU, locks the RAM, and inserts the GPU into a PCIe slot. That last one is the crazy part. A GPU slot gives you almost no room for error. Sometimes humans can't assemble! If the angle is slightly wrong, the card catches the rim and jams. CoStream from Harvard, Stanford, MIT, and Columbia handles this by not betting everything on one giant robot policy. Instead, it composes simple behaviors: 🔵 language vision to understand the task 🔵 a video world model to imagine the motion 🔵 tactile feedback to detect slip and misalignment 🔵 force control to push without destroying the hardware The important part is not just that the robot knows where the GPU should go. It can feel when reality disagrees with the plan. That is why the assembly demos are interesting.