DexForce

Our Mission is to develop a sustainable roadmap towards Physical AGI.

[DexForce @DexForceAI]

Why does a robot need 8B parameters to pick up a cup? The hidden cost in robot learning is translation, not computation.

Introducing UNITAS, the first task-native 3D world action model for embodied manipulation.

Most world models burn capacity translating 2D images into physics instead of understanding the world directly. Simulation logs, real robot trajectories, and human demos sit in separate silos that no single model can absorb. UNITAS models observation, action, and scene change natively in one 3D coordinate system. Every data source becomes one language, points moving through 3D space.

The result. 1.7B parameters beating models up to 8B. 99.8% on LIBERO. 93% on RoboTwin 2.0 bimanual tasks. 85% average success on real robots.
Less compute. More data sources working together. Real-world performance.

Paper: https://t.co/phGEgHdi70
Code: https://t.co/6c3MSwzta0

[DexForce @DexForceAI]

Round three looks exactly like round one.

The red block lifts and returns to the same spot three times. Each round ends on a near-identical frame. A single-frame policy sees the same picture every time.

TaskAnchor keeps the count. Round one finishes, round two follows, round three ends the task. Memory across repeated stages, continuous execution, and a clean stop on time.

Figure 7 isolates this test from the repeated lifting task in Figure 5b. Same scene, different state.

#EmbodiedAI #RobotLearning #Manipulation

[DexForce @DexForceAI]

Point. Put back. Walk away. The table looks untouched.

Three trials, three prompt objects. A person lifts a doll, a chip bag, or a bottle, uses the object to point at a target, then returns the object and withdraws the hand. The layout comes back to nearly the same arrangement. The final frame carries no cue about the instruction.

TaskAnchor picks the object the person just indicated. A policy reading only the current frame cannot tell the three trials apart.

This is the failure mode behind long-horizon manipulation, and the reason for the numbers in the TaskAnchor release. Two frames look identical. The correct next action differs. Only the past separates those cases.

79% on RMBench against 14% for the original π0.5.

GIF below. Paper https://t.co/aoZQqzTqYb

[DexForce @DexForceAI]

Long-horizon robots fail at the same frame, not at the same skill. 🧠
79% average success on RMBench. 5.6x the original π0.5.
Robots have learned many manipulation skills. Long-horizon tasks add a harder question. Two identical frames can demand different next moves. Only the past separates those cases.

TaskAnchor gives a pretrained policy history memory and stage awareness. CA-TTT keeps key past cues in view. A Task-State Coordinate marks the current stage. The model calls the right skill at the right moment.

No extra planner. No architecture change. 27M parameters, 2ms per action chunk.

RMBench 79% against 14%.
RoboMemArena 50.0% against 21.5%.
Real robot 75 trials, 75 successes.

Long-horizon manipulation may need less new skill and more memory.
Project page https://t.co/l0VFLeKfo3
Paper https://t.co/tBejXKi8U0

#EmbodiedAI #RobotLearning #Manipulation

[DexForce @DexForceAI]

Embodied AI meets industrial automation at SPS 2026 🎉

DexForce is heading to Nuremberg this November, bringing the W1 Pro wheeled humanoid robot and production-proven industrial 3D vision solutions to the heart of the automation industry.
From machine tending to bin picking, these solutions are already running in production, not sitting in labs.

Meet DexForce at [Hall 4A, Booth 467], November 24 to 26, Nuremberg Exhibition Centre.
See embodied intelligence at work.
#SPS2026 #EmbodiedAI #HumanoidRobot #IndustrialAutomation #3DVision #MachineTending

[DexForce @DexForceAI]

Thank you @Reuters for featuring DexForce!
We’re using simulation to help robots learn more efficiently and turn virtual training into real-world capability.
We’re building toward a future where Physical AI enables robots to understand, adapt to, and act in the real world.

[Reuters @Reuters]

At the offices of Chinese startup DexForce, a screen shows a virtual robot placing objects on a table. It repeats the task as the angle of its arms, the colour of the table and the objects themselves change.

DexForce is betting that such simulated training can help solve one of the humanoid-robot industry's biggest challenges: teaching machines to perform a wide range of real-world tasks without human assistance https://t.co/s08RVIOP5f

[DexForce @DexForceAI]

One brain, many jobs.

A coffee bar in a mall. A popcorn counter in a park. A phone inspection cell on a production line. Same embodied brain underneath, planning and generalizing across all of them.W1 Pro ships with that brain.

New task, same unit, new fluency.
2000+ embodied intelligence products shipped across 50+ applications.

[DexForce @DexForceAI]

Fair reading. On LIBERO the 3D-native route reaches 99.8% average success at 1.7B parameters, and both RoboTwin 2.0 dual-arm settings clear 92%. The cup also shows where the next problem sits, since flexible items and cluttered bins punish a depth guess harder than a rigid body does.

[DexForce @DexForceAI]

Why does a robot need 8B parameters to pick up a cup? The hidden cost in robot learning is translation, not computation.

Introducing UNITAS, the first task-native 3D world action model for embodied manipulation.

Most world models burn capacity translating 2D images into physics instead of understanding the world directly. Simulation logs, real robot trajectories, and human demos sit in separate silos that no single model can absorb. UNITAS models observation, action, and scene change natively in one 3D coordinate system. Every data source becomes one language, points moving through 3D space.

The result. 1.7B parameters beating models up to 8B. 99.8% on LIBERO. 93% on RoboTwin 2.0 bimanual tasks. 85% average success on real robots.
Less compute. More data sources working together. Real-world performance.

Paper: https://t.co/phGEgHdi70
Code: https://t.co/6c3MSwzta0

[DexForce @DexForceAI]

DexForce at Hong Kong Electronics Fair — RoboPark, Booth RH-07.

We brought a robot that plays Chinese chess. It hasn't stopped since the doors opened.

Three things worth watching at Booth RH-07:
- Long-duration stability.
Running continuously through the busiest hall at this show — no resets, no babysitting, no "please wait, rebooting."
- Precise recognition of the board.
Every piece, every square, read accurately — including everything that moved since the last turn.
- Fast interaction.
Perceive, reason, act in one closed loop. It responds to the board's actual state in real time, not a prerecorded sequence.

Day 1, and an unexpected one:
more serious chess players at this show than we expected. Players lined up to try to beat it, and the robot kept giving them a game.

Three more days — Oct 14–16. Come by RH-07 and bring your best opening.

[DexForce @DexForceAI]

110,000+ visits. 500+ orders per robot per day. One golden week.
This National Day, DexForce ran 30+ robotic service stations across China. Humanoid robots served ice cream, popcorn, coffee, and drinks. Sites spanned tourism spots, theme parks, and shopping complexes. Visits reached 110,000+ during the holiday. A single robot served 500+ orders per day on average. The flagship station earned RMB 150K. Commercial deployments have logged 25,000+ hours of uninterrupted operation to date.
Robot value is shifting. Single-task execution gives way to continuous operation. Technology validation gives way to business validation. One-off projects gives way to scale replication. DexForce keeps proving the path.
#EmbodiedAI #HumanoidRobot #Robotics

[DexForce @DexForceAI]

Can one robot work a retail counter and a factory floor?

Popcorn. Ice cream. Machine tending. One embodied brain.

Spec sheet in DM. #EmbodiedAI #HumanoidRobot

[DexForce @DexForceAI]

Thanks to @asia_bits for visiting the DexForce booth at WRC 2026.
Thomas watched W1 Pro handle live serving and retail-station tasks on the show floor, and spent time discussing the Sim2Real pipeline behind the robot. W1 Pro trains on 100% synthetic data through the self developed AI engine, building the vision-language-action model that runs everyday commercial work.
Drop by to see W1 Pro serving, stocking and restocking in real time, and to talk technical details with the team.

[DexForce @DexForceAI]

Thank you for your interest in DexForce and for the detailed questions. The answers below address each point in turn, and hopefully clarify the full picture.

1. SKU Changeover Time

The post-training data for models used on production lines comes 100% from simulation environments. Data generation and model adaptation for a new task can be completed within hours. In practice, a new SKU goes from data generation to production-ready adaptation in a matter of hours, without the need for real production data or extended downtime.

2. Self-Correction and Error Recovery

Error recovery is often the top concern on real production lines, so this part covers the topic in more detail.

Robustness against unexpected line interference and the robot's self-correction after errors are treated as two separate metrics. Unexpected situations are classified into three types for recovery-rate tracking.

a. Target displacement recovery. Perception or target changes, such as a part shifting position on a conveyor.
b. Grasp failure recovery. Operation failures, such as an object slipping during grasping.
c. Line condition anomalies. Abnormal line states, such as material build-up.
Based on testing, recovery rates for types a and b exceed 92%. Human assistance currently focuses mainly on type c, where line-level conditions call for additional monitoring.

If your production line involves different recovery scenarios, feel free to reach out for a more detailed discussion.

[DexForce @DexForceAI]

Line speed is the easy part. Keeping up is the hard part.

Powered by the embodied brain, the robot tracks moving targets, assembles with precision, inspects every unit, and self-corrects on the fly. One closed loop, from moving box to finished pack.

DexForce | Embodied AI for Smart Manufacturing
#EmbodiedAI #PackagingAutomation #Robotics

[DexForce @DexForceAI]

DexForce at WRC 2026: On‑site Highlights

[DexForce @DexForceAI]

Vielen Dank für den Besuch in unserem Büro in Shenzhen und diesen tollen Beitrag!
Mit unserem Sim2Real‑Workflow überbrücken wir die Lücke zwischen Simulation und realer Roboter‑Einsatzpraxis. Wir arbeiten weiter an generalisierter Embodied‑KI für anspruchsvolle Handhabungsaufgaben in der Industrie.
Wir freuen uns auf die weitere Zusammenarbeit mit deutschen Industriepartnern.

Great write‑up from our on‑site visit, thank you for coming to see our team in Shenzhen!

Our core focus is closing the gap between simulation and real‑world deployment with Sim2Real / Real2Sim2Real workflow. We keep pushing generalized embodied AI to solve tricky industrial manipulation tasks, from automotive manufacturing to delicate object handling.

Excited about our ongoing collaboration with German industry partners. More real‑world use‑cases coming soon.

[Leonard M. Schmedding @derleomartin]

Diese Humanoiden Roboter arbeiten LÄNGST in der Fabrik und kaum einer kennt sie!

Ich bin zu Besuch bei @DexForceAI in Shenzhen, einem der aufstrebendsten KI- und Robotik-Start-ups Chinas.

Hier entsteht nicht nur die Hardware, sondern das komplette System aus Simulation, Robotikmodellen und Software.

DexForce hat den Unicorn-Status bereits erreicht und mehr als 1.000 Projekte weltweit in über 50 Branchen.

Spezialisierte Robotermodelle entstehen hier in nur 6 bis 8 Stunden. Auch die erste Zusammenarbeit mit einem deutschen Automobilkonzern läuft bereits!

Das habe ich vor Ort gesehen:

- Sim2Real ersetzt 3 bis 6 Monate reale Datensammlung
- Trainingswelten entstehen aus Text, Bildern und 3D
- Licht, Positionen und Bauteile werden digital variiert
- Ein Robotermodell entsteht in nur 6 bis 8 Stunden
- Real2Sim2Real spielt jeden Fehler zurück ins Training
- Ein World Action Model sagt Störungen voraus und korrigiert Bewegungen
- Schach mit DeepSeek und eigenen Aktionsmodellen
- Die schwierigste Aufgabe: glatte Kaffeekapseln greifen und bei Fehlern nachfassen

„Die Roboterhand muss eine glatte Kaffeekapsel greifen, erkennen, ob der Griff funktioniert hat, und es bei einem Fehler selbstständig erneut versuchen."

„Der Kaffee schmeckt wirklich gut. Und im Gegensatz zu manchen Menschen lernt der Roboter sogar aus seinen Fehlern.“

„Diese Roboter stehen nicht mehr im Labor. Sie arbeiten längst bei BYD, in der Chemieindustrie und inzwischen auch für einen deutschen Automobilkonzern.“

[DexForce @DexForceAI]

Sim2Real has left the lab. The hard part now is commercial scale.
Yueci Deng, Chief Engine Architect at DexForce, joins Rachid Daou, Joanne Zhiyan Wu, and Peter Haas on September 3 to talk about turning promising models into deployable systems.
📅 September 3 | 9:00 AM ET | 3:00 PM CEST | 9:00 PM UTC+8
Register：
https://t.co/t297PejMQU

[DexForce @DexForceAI]

Seventeen children from Shanxi's old revolutionary base regions walked into DexForce last week, and got a first-hand look at embodied AI. 💌

The children shook hands with humanoid robots, watched a robot brew coffee, and challenged a robot to a chess match. No glass cases, no distant demos. Machines moving, working, and playing, within arm's reach.

For DexForce, real value lies in making technology feel approachable. A single visit can plant the first seed of scientific curiosity.
Each child left with a gift and, hopefully, a new question to explore. Technology advances industries, and carries warmth along the way.
DexForce stays committed to bringing the light of technology to growing dreams.🌱

[DexForce @DexForceAI]

🔎Discover how DexForce coordinates a humanoid and a cobot for fully autonomous loading and unloading of sheet metal parts.🤖
The humanoid robot locates the bin with vision and completes pick and dump actions with ±0.5 mm arm repeatability, while laser SLAM navigation keeps it running a full 8-hour shift on a single one-hour charge. ⚙️ Guided by AI and 3D vision, the cobot arm delivers 99.99% bin-picking accuracy at one second per item, mastering even reflective, diffuse and transparent surfaces. 📈 CAD-pretrained models enable deployment in just 8 hours, cutting integration costs and speeding time to production.
Built for automotive 🚗, sheet metal fabrication, machining and beyond.
#EmbodiedAI #BinPicking #MachineVision #HumanoidRobot #SmartManufacturing

[DexForce @DexForceAI]

DexForce brings zero-shot adaptive grasping and intelligent obstacle avoidance to the future of retail automation. 🚀

Say goodbye to rigid, preset checkout systems. Our latest robot doesn't just scan—it sees, thinks, and adapts to real-world chaos in real time.
🔹 True Flexibility: Picks completely untrained items directly from bare tables—no fixed trays or preset positions needed.
🔹 Seamless Human Collaboration: Flawlessly accepts hand-offs, instantly adjusting the item to align the barcode for quick scanning.
🔹 Unstoppable Workflow: Senses and dodges obstacles on the fly, driving the end-to-end checkout process completely autonomously.
Bring next-level adaptability to your operations.

See the tech in action below!
Our Website: https://t.co/tJtg8LYNsk