Building generalist robots for real-world deployment Open models, benchmarks & uncut demos. WALL-OSS-0.5 · WALL-WM · XRZero-G0 ↓ x2robot.com/en/research
Global
- Tweets 287
- Following 178
- Followers 1,959
- Likes 278
Pinned Tweet
What if a dexterous robot could learn a 20+ step chemistry experiment without a single on-robot training demo?
Meet TwinDEX: a pair of co-designed, three-finger, nine-DoF dexterous manipulation interface: one wearable for data collection, one for robot deployment.
The twinned design shares identical kinematics, contact surfaces, visual appearance, and sensors across collection and deployment — keeping observations and actions aligned end to end.
Trained from scratch on only a few hundred wearable demonstrations - with zero on-robot training or intervention data - TwinDEX completed a standardized chemistry experiment involving tool switches, fine force control, and bimanual coordination.
Robot-free data showed comparable learning efficiency on the multi-task evaluation, TwinDEX delivered 5.3 times effective throughput than on-robot teleoperation.
TwinDEX demonstrates that high-quality robot-free data can fully substitute for on-robot teleoperation data on challenging dexterous tasks — removing the dependency on real-robot hardware that has been the central bottleneck to scaling dexterous manipulation data.
This was the proof-it phase. Now comes scale: what emerges at tens of thousands, or millions, of episodes?
Watch the demo and read the technical blog: x2robot.com/en/pages/twindex
#TwinDEX #Robotics #EmbodiedAI #DexterousManipulation
65
57
380
1,346,019
Happy to introduce X-Planner: Event-Structured Task Planning for Embodied Intelligence — a front-end that treats semantic events as planning units, exposes discrete + latent interfaces via Staircase Decoding, and outperforms baselines on real robots.
To address long-horizon robotic manipulation tasks, we propose X-Planner, an embodied long-horizon task planner:
It decomposes high-level natural language instructions into event-level subtasks, and can also output continuous implicit Chain-of-Thought (CoT). It directly interfaces with downstream VLA or WAM models to operate robots.
GitHub: github.com/X-Square-Robot/Xp…
Checkpoint: huggingface.co/x-square-robo…
Benchmark: huggingface.co/datasets/x-sq…
2
344
Join the X Square Robot IROS 2026 After Party in Pittsburgh! 🤖🎉
Meet fellow researchers and builders, exchange ideas, and celebrate embodied AI together.
Scan the QR code to register. See you there!
X Square Robot is coming to Pittsburgh for IROS 2026!
From September 27 to October 1, the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) will bring the global robotics community together to share groundbreaking research, exchange ideas, and shape the future of robotics and intelligent systems.
Stop by Booth 727 from September 27–30 to meet the X Square Robot team and discover how we are advancing embodied AI through scalable physical experience, simulation, physical world models, and predictive embodied action.
We’re excited to meet researchers, builders, and everyone passionate about embodied intelligence and its real-world applications.
Schedule a meeting with our on-site team: contact@x2robot.com
See you in Pittsburgh!
#IROS2026 #EmbodiedAI #PhysicalAI #Robotics @ieeeiros
4
480
We're proud to see our QUANTA X1 Pro wheeled bimanual robot deployed at Lululemon’s Wuhan Distribution Center, which officially opened on September 16.
Developed by @lululemon in partnership with @SFlogistic , the center uses automated equipment and end-to-end RFID-enabled processes. On the logistics line, QUANTA X1 Pro handles and organizes parcels as part of daily operations.
Following our live dual-arm logistics demonstration at WRC, QUANTA X1 Pro is now working in a real warehouse environment.
#EmbodiedAI #Robotics #SmartLogistics
Made with AI
3
8
63
25,703
X Square Robot is coming to Pittsburgh for IROS 2026!
From September 27 to October 1, the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) will bring the global robotics community together to share groundbreaking research, exchange ideas, and shape the future of robotics and intelligent systems.
Stop by Booth 727 from September 27–30 to meet the X Square Robot team and discover how we are advancing embodied AI through scalable physical experience, simulation, physical world models, and predictive embodied action.
We’re excited to meet researchers, builders, and everyone passionate about embodied intelligence and its real-world applications.
Schedule a meeting with our on-site team: contact@x2robot.com
See you in Pittsburgh!
#IROS2026 #EmbodiedAI #PhysicalAI #Robotics @ieeeiros
1
15
1,080
Excited to join Saturday Robotics × IROS 2026 in Pittsburgh! 🤖
We will present X2Real, our extensive simulation benchmark for evaluating real-world generalist robot policies—featuring 44 hierarchical, long-horizon tasks across 10 capability dimensions.
We look forward to sharing our latest work and connecting with researchers and builders advancing robot learning, simulation-to-real transfer, and embodied AI.
📍 Pittsburgh
📅 September 28, 2026
👉🏻 luma.com/tzbw7n61
See you there!
🍾🍲 Saturday Robotics x IROS 2026 — Robotics Research Night
👉🏻 luma.com/tzbw7n61
We’re bringing a high-signal evening of robotics research to Pittsburgh on September 28.
After a full day at IROS, we’ll bring together researchers, engineers, founders, students, and investors for technical discussions, networking, and a series of ~10-minute lightning talks.
Tentative preview of the current lineup:
🤖 1. PAC-MAN: Perception-Aware CBF-RL for Whole-Body Safety in Humanoid Dodgeball
Gary Yang @lzyang2000 (@Caltech)
Perception-aware reinforcement learning + Control Barrier Functions for whole-body humanoid safety. Demonstrated on a Unitree G1, with 19/20 successful dodges and zero falls in real-world experiments.
🧠 2. How In-Context Learning Is Reshaping Robot Learning Data at Scale
AaronLi (@RhodaAI)
Exploring how in-context learning can change the way we think about robot learning data, scaling, and generalization.
🧪 3. X2Real: An eXtensive Simulation Benchmark for Real-World Generalist Policies
Liangwang Ruan (@XSquareRobot)
A new simulation benchmark built around faithfulness, diversity, and fairness, with 44 hierarchical long-horizon tasks across 10 capability dimensions and a reported 0.84 simulation-to-real correlation.
🦾 4. Rethinking Generalist Robotic Manipulation: Architecture, Data and Inference for Real-World Deployment
Peiyan Li (Chinese Academy of Sciences, @CAS__Science)
3D VLA architectures, memory augmentation, ego/UMI human priors, large-scale robot pretraining, and inference-time contextual learning for deployable generalist manipulation.
🎯 5. HiRE: Hindsight Reward Editing for Policy Finetuning
Haoyi Niu @t641769919 (@UCBerkeley)
Accepted at CoRL 2026. A training-free approach to reward editing that uses successful and failed trajectories to identify “trap states” and provide denser, control-aware feedback for RL.
🔥 6. Lightning Talk — Open Slot
We’re opening one additional slot for a technically deep research talk, new project, frontier paper, demo, open problem, or startup technical insight.
10 minutes. A few slides. One sharp technical idea. No fluff.
Topics include World Models, Physical AI, Humanoids, VLAs, Robot Foundation Models, Manipulation, RL, Simulation & Sim-to-Real, Spatial Intelligence, Computer Vision, and Embodied AI.
📍 Pittsburgh
📅 September 28, 2026
🕠 5:30–9:30 PM
🍾 Networking + Technical Talks + Research Discussion
📩 junfanzhu98@gmail.com
See you in Pittsburgh. 🤖
#IROS2026 #Robotics #PhysicalAI #RobotLearning #WorldModels #HumanoidRobotics #VLA #EmbodiedAI #RobotFoundationModels
2
8
1,266
X Square Robot retweeted
🍾🍲 Saturday Robotics x IROS 2026 — Robotics Research Night
👉🏻 luma.com/tzbw7n61
We’re bringing a high-signal evening of robotics research to Pittsburgh on September 28.
After a full day at IROS, we’ll bring together researchers, engineers, founders, students, and investors for technical discussions, networking, and a series of ~10-minute lightning talks.
Tentative preview of the current lineup:
🤖 1. PAC-MAN: Perception-Aware CBF-RL for Whole-Body Safety in Humanoid Dodgeball
Gary Yang @lzyang2000 (@Caltech)
Perception-aware reinforcement learning + Control Barrier Functions for whole-body humanoid safety. Demonstrated on a Unitree G1, with 19/20 successful dodges and zero falls in real-world experiments.
🧠 2. How In-Context Learning Is Reshaping Robot Learning Data at Scale
AaronLi (@RhodaAI)
Exploring how in-context learning can change the way we think about robot learning data, scaling, and generalization.
🧪 3. X2Real: An eXtensive Simulation Benchmark for Real-World Generalist Policies
Liangwang Ruan (@XSquareRobot)
A new simulation benchmark built around faithfulness, diversity, and fairness, with 44 hierarchical long-horizon tasks across 10 capability dimensions and a reported 0.84 simulation-to-real correlation.
🦾 4. Rethinking Generalist Robotic Manipulation: Architecture, Data and Inference for Real-World Deployment
Peiyan Li (Chinese Academy of Sciences, @CAS__Science)
3D VLA architectures, memory augmentation, ego/UMI human priors, large-scale robot pretraining, and inference-time contextual learning for deployable generalist manipulation.
🎯 5. HiRE: Hindsight Reward Editing for Policy Finetuning
Haoyi Niu @t641769919 (@UCBerkeley)
Accepted at CoRL 2026. A training-free approach to reward editing that uses successful and failed trajectories to identify “trap states” and provide denser, control-aware feedback for RL.
🔥 6. Lightning Talk — Open Slot
We’re opening one additional slot for a technically deep research talk, new project, frontier paper, demo, open problem, or startup technical insight.
10 minutes. A few slides. One sharp technical idea. No fluff.
Topics include World Models, Physical AI, Humanoids, VLAs, Robot Foundation Models, Manipulation, RL, Simulation & Sim-to-Real, Spatial Intelligence, Computer Vision, and Embodied AI.
📍 Pittsburgh
📅 September 28, 2026
🕠 5:30–9:30 PM
🍾 Networking + Technical Talks + Research Discussion
📩 junfanzhu98@gmail.com
See you in Pittsburgh. 🤖
#IROS2026 #Robotics #PhysicalAI #RobotLearning #WorldModels #HumanoidRobotics #VLA #EmbodiedAI #RobotFoundationModels
5
14
2,217
Most "streaming" TTS systems still wait for a complete sentence before speaking.
X2Streaming-TTS removes that wait. It consumes text tokens as they arrive and generates speech with strict zero lookahead.
The challenge: once speech is played, it cannot be revised.
Take “He finished 3…”—should “3” become “three” in “3 laps,” or “third” in “3rd place”? Speaking too early risks an error; waiting defeats the purpose of streaming.
X2Streaming-TTS addresses this with:
1️⃣ Causal commitment
Ambiguous numbers, units, and symbols are held until their pronunciation becomes clear. Segments are closed using both punctuation and acoustic capacity.
2️⃣ Speech-state inheritance
Waveform-decoder state and acoustic history are carried across segments, preserving pitch, timbre, and continuity.
Results:
⚡ 15.8 ms median TTFT for one request
⚡ Under 120 ms at 64 concurrent requests
🎯 Lowest recognition error in 6/8 evaluated streaming conditions
🔢 0% CER on numeric and streaming-ambiguity tests
🗣️ Quality comparable to evaluated offline baselines
Low-latency voice AI is not just about speaking faster—it is about knowing when it is safe to speak.
Paper: arxiv.org/abs/2608.18661
Code: github.com/X-Square-Robot/X2…
#TTS #VoiceAI
2
11
78
3,044
X Square Robot retweeted
Dexterous robot learning has a data scaling problem.
Chinese startup X Square Robot's TwinDEX is taking an interesting approach by pairing a wearable manipulation interface with a structurally matched dexterous hand for the X Square Robot.
Instead of collecting demonstrations directly on the robot, humans can perform the tasks while the wearable system captures the motion, contact, visual, and timing information needed for training.
The hardware is designed to stay aligned across data collection and robot execution, reducing the embodiment gap that can make robot-free demonstrations difficult to transfer.
The reported results are compelling:
→ 20+ sub-actions in a standardized chemistry workflow
→ Policies trained from scratch using robot-free data only
→ Comparable learning efficiency to on-robot teleoperation data
TwinDEX is essentially treating the collector, robot hardware, data, and policy as one system — a promising direction for scaling real-world dexterous robot data.
23
22
89
15,093
Safer robots start with better world models. 🤖
Meet Shalfun Li, World Model Tech Lead at X Square Robot, at the Safe World Models workshop during #ECCV2026.
📅 Sept. 9 · Morning (CEST)
📍 Malmömässan D1
🌐 Half-day hybrid event
Let’s connect in Malmö!
trustworthy-world-models.git…
Made with AI
2
342
X Square Robot retweeted
Robot learns 20+ step chemistry experiment without a single on-robot training demo.
Using only a few hundred wearable demos, with zero real-robot training, TwinDex by X Square Robot (自变量机器人) mastered a complex 20+ step chemistry experiment involving tool switches, fine force control, and bimanual coordination.
The old way you had to either physically grab the robot's metal arm through every motion, or wear a controller to drive it, keeping the robot powered on and mirroring your every move for the entire multi-step procedure.
Now with the new way you just wear a glove that matches the robot's hand, perform the task at your own pace while the robot stays turned off or works on other things, then replay the recording to the robot.
It frees up the hardware and slashes training time by 5x.
1
2
206
X Square Robot retweeted
Dexterous robot data is expensive to collect. TwinDEX may have found a way to make it much more scalable.
The system pairs a wearable manipulation interface with a structurally matched dexterous robotic hand for the X Square Robot, allowing humans to demonstrate tasks without occupying the robot itself.
The interesting part is that the wearable interface and the robot hand are designed as two sides of the same learning system, aligning kinematics, contact mechanics, vision, sensing, and timing.
In reported experiments, policies trained from scratch on robot-free data achieved learning efficiency comparable to on-robot teleoperation data.
The collection throughput?
More than 5× higher than on-robot teleoperation.
For physical AI, the ability to collect more robot-relevant data without putting a robot in the loop could be a big deal.
10
26
92
55,495
X Square Robot retweeted
Robot intelligence is improving fast, but there’s still a major bottleneck: data.
Dexterous manipulation requires huge amounts of real-world demonstrations. But collecting that data directly on robots can be expensive, difficult to scale, and heavily dependent on the hardware.
TwinDEX is exploring a different approach.
Its system combines a wearable manipulation interface with a structurally matched dexterous hand for the X Square Robot, allowing humans to demonstrate complex actions without directly operating the robot.
The system captures the motion, contact, and visual information needed for training while keeping the data closely aligned with the robot that will ultimately execute the task.
Early results are promising across contact-rich tasks, including tool use, fine manipulation, and coordinated actions.
If we want physical AI to scale, better ways of collecting and transferring real-world data may be just as important as better models.
36
85
114
21,293
X Square Robot retweeted
One challenge in robot learning is often overlooked:
Data collected with one embodiment does not necessarily transfer well to another.
A human hand, a parallel gripper, and a dexterous robot such as the X Square Robot all have different contact mechanics, kinematics, and action spaces. Simply retargeting motion can lose the fine-grained details that make a manipulation task work.
TwinDEX takes a different approach.
It co-designs a wearable manipulation interface with a structurally matched dexterous hand for the X Square Robot, aligning motion structure, contact surfaces, sensing, vision, and timing.
The result is robot-free demonstration data that remains closely aligned with the robot being trained.
In one reported experiment, a policy trained only on wearable demonstrations completed a chemistry workflow with 20+ sub-actions.
No on-robot teleoperation data was used for training.
An interesting example of hardware, data, and policy co-design for scaling dexterous manipulation.
8
37
47
22,516
X Square Robot retweeted
This is huge news for dexterity manipulation, sometimes it takes thinking outside the box to come to the conclusion that actually maybe a turtle hand is enough, 3 fingers not 5.
Also impressive that no training was done on robot, but rather a co-designed wearable for data collection with a human.
In the next 12-24 months we will see an ever accelerating arena for robotics, not just humanoids, but every single vertical within the field.
This is truly a pivotal moment in time, and we are just getting started.
What if a dexterous robot could learn a 20+ step chemistry experiment without a single on-robot training demo?
Meet TwinDEX: a pair of co-designed, three-finger, nine-DoF dexterous manipulation interface: one wearable for data collection, one for robot deployment.
The twinned design shares identical kinematics, contact surfaces, visual appearance, and sensors across collection and deployment — keeping observations and actions aligned end to end.
Trained from scratch on only a few hundred wearable demonstrations - with zero on-robot training or intervention data - TwinDEX completed a standardized chemistry experiment involving tool switches, fine force control, and bimanual coordination.
Robot-free data showed comparable learning efficiency on the multi-task evaluation, TwinDEX delivered 5.3 times effective throughput than on-robot teleoperation.
TwinDEX demonstrates that high-quality robot-free data can fully substitute for on-robot teleoperation data on challenging dexterous tasks — removing the dependency on real-robot hardware that has been the central bottleneck to scaling dexterous manipulation data.
This was the proof-it phase. Now comes scale: what emerges at tens of thousands, or millions, of episodes?
Watch the demo and read the technical blog: x2robot.com/en/pages/twindex
#TwinDEX #Robotics #EmbodiedAI #DexterousManipulation
5
3
35
7,964
It’s like everyday is a new breakthrough in robotics.
X Square just benchmarked a three finger robot hand on tasks like twisting bottle caps, opening books, operating syringes, and even a full 24-step chemistry experiment.
The setup is called TwinDEX. Instead of teleoperating the robot directly, they built a wearable hand that matches the robot hand’s geometry, sensors, and contact surfaces, so humans can demonstrate tasks naturally and the data transfers much more cleanly to the robot. X Square says it reached up to 5.3x the effective data-collection throughput of on-robot teleoperation.
The other part I like is that they actually tested how many fingers and DoF you really need. For their current task set, three fingers with 7 active DoF gave the best balance of dexterity, cost, complexity, and reliability.
For cheaper humanoids, maybe companies that want to deliver cost effectively shouldn’t ask “how do we copy the human hand exactly?” but “where is the best trade-off?” (This is probably obvious to them) however what’s not is the right balance that happens to be benchmarked. A strong thumb plus two opposing fingers might capture most useful manipulation while keeping hardware far cheaper, simpler, and more reliable.
What if a dexterous robot could learn a 20+ step chemistry experiment without a single on-robot training demo?
Meet TwinDEX: a pair of co-designed, three-finger, nine-DoF dexterous manipulation interface: one wearable for data collection, one for robot deployment.
The twinned design shares identical kinematics, contact surfaces, visual appearance, and sensors across collection and deployment — keeping observations and actions aligned end to end.
Trained from scratch on only a few hundred wearable demonstrations - with zero on-robot training or intervention data - TwinDEX completed a standardized chemistry experiment involving tool switches, fine force control, and bimanual coordination.
Robot-free data showed comparable learning efficiency on the multi-task evaluation, TwinDEX delivered 5.3 times effective throughput than on-robot teleoperation.
TwinDEX demonstrates that high-quality robot-free data can fully substitute for on-robot teleoperation data on challenging dexterous tasks — removing the dependency on real-robot hardware that has been the central bottleneck to scaling dexterous manipulation data.
This was the proof-it phase. Now comes scale: what emerges at tens of thousands, or millions, of episodes?
Watch the demo and read the technical blog: x2robot.com/en/pages/twindex
#TwinDEX #Robotics #EmbodiedAI #DexterousManipulation
3
6
83
10,331
X Square Robot retweeted
Three fingers can get a lot done.
Latest release from @XSquareRobot: TwinDEX, a robot free dexterous manipulation system pairing a wearable 3 finger, 9 DoF collector with a closely matched deployment hand.
The pair aligns kinematics, contact geometry, sensing and timing, while synchronizing multiview RGB, 6 DoF wrist pose, finger states and fingertip tactile data.
They report up to 5.3x the effective throughput of on robot teleoperation and say a policy trained from only a few hundred robot free episodes completed a 24 step chemistry experiment autonomously in one run.
X Square Robot
1
4
24
1,720