Robot-Free Data, Physics-Aware Video Models and Agent Sandboxes Advance

According to Leiphone, Chinese robotics company 自变量机器人 released TwinDEX, a system built around wearable, robot-free data collection and a matching robot end effector. In a demo, a robot performed 24 sub-actions in one continuous take, including opening a bottle and taking a sample, using a dropper, guiding liquid with a glass rod, mixing and shaking. The training data consisted of only several hundred robot-free examples, with no real-machine data. The company said tests show robot-free data can almost 100 percent replace real-machine teleoperation data for training. TwinDEX uses a three-finger dexterous hand with nine degrees of freedom, seven of them active. In earlier data collection schemes, robot-free data mixed with real-machine data at a 10:1 ratio was comparable to full real-machine training. In new multi-task tests, pure robot-free policies and real-machine policies improved almost in parallel as data increased and reached similar levels. The company also said data collection efficiency for relevant tasks was 5.3 times that of real-machine teleoperation.

According to MarkTechPost, researchers from NVIDIA, MIT and the University of Oxford introduced Physis-Lang, a framework that treats physical language as a shared and optimizable representation. The same text drives data curation, model training and inference. On a public Physics-IQ Verified leaderboard snapshot dated September 29, 2026, Physis-Lang on Cosmos3-Super ranked first at 48.2 ± 1.4, while Cosmos3-Nano ranked second at 43.3 ± 1.5. The framework adds a physics reasoning field to captions and writes scene-specific negative prompts describing implausible outcomes. A captioner writes captions for a fixed 20-video development set with 273 human-verified assertions, and Gemini-3.1-Pro acts as a physics-aware critic. Caption F1 rose from 78.64 at iteration 1 to 87.82 at iteration 9, after dropping to 76.28 at iteration 2. The final training set held 183,000 videos, including 71,000 filtered from WISA-80K and 112,000 retrieved clips. Retrieval alone added 3.01 points on average across three benchmarks. Compared with Google's Veo 3.1, Physis-Lang on Cosmos3-Nano scored 71.04 versus 65.63 on PhyGenBench, 43.41 versus 34.99 on Physics-IQ Verified, 69.90 versus 69.24 on PhyGround, and 68.02 versus 68.87 on VideoPhy-2 full set, 62.36 versus 58.43 on the Hard split. The team distilled the pipeline into two Qwen3-VL-4B-Instruct models; on Wan2.1-14B, PhysThinker-C kept +6.76 at about $0.12K in API cost, versus +7.05 at about $24.12K for the commercial pipeline, while a fully local setup cost $0 and added +4.76.

According to a Zhihu article republished by 量子位, DeepSeek published a technical article on DSec, or DeepSeek Elastic Compute, the sandbox infrastructure supporting all training, evaluation and data preprocessing for DeepSeek-V4. The arXiv paper was released with Tsinghua University and has more than 130 authors, including Liang Wenfeng. DSec supports FnCall, Container, MicroVM and Full VM backends through a unified Python SDK called libdsec. In one week of 2026 production data, the container backend used 11,266 base images, 102,171 workspaces and hundreds of toolkits. DSec stores images, workspaces and toolkits in EROFS format and combines layers with OverlayFS. Analysis found that runtime access touched only 4.2 to 13.3 percent of total image size. In an experiment creating 8,192 containers, on-demand loading cut completion time from more than 60 minutes to about 35 minutes, a 1.71 times speedup, and reduced disk writes by about 57 percent. Changing workspace supply from unpacking tar.gz files to mounting EROFS layers cut time from 79 minutes to 45 minutes and reduced total disk writes to about 1/5.5 of the original. DSec reported an oversell rate above 50 times in production. Enabling virtio-pmem and DAX alone reduced peak host memory use by 40.2 percent; DAMON and balloon idle-page reporting reduced cumulative host memory consumption by 21.2 percent. CPU scheduling optimization cut the latency increase for latency-sensitive tasks from 45.2 percent to 17.3 percent when other tasks used 50 percent of node CPU capacity. From DeepSeek-V4.1, rollout execution was moved to DSec sandboxes split into an agent sandbox and a worker container, both outside the preemptible GPU resource pool. A DSec shard has about 160 servers, 30,000 CPU cores and 250 TB memory, serving about 3 million sandboxes per day, with peak concurrency above 380,000 and more than 5,000 sandbox creations per second. Multiple shards support millions of concurrent sandboxes. DSec uses AppArmor and eBPF for access control.

According to Leiphone, Yang Sicheng, founder of 源升智能, said dexterous hands must first compete on performance before embodied model capabilities break through. Yang studied dexterous manipulation at Tsinghua University and led the dexterous hand project at Tencent Robotics X before founding the company in 2024. Its Apex Hand uses a hybrid drive scheme, mainly tendon-driven with direct drive and linkage transmission, and has 21 degrees of freedom. It passes the Kapandji hand dexterity test and has a maximum single-hand active load of 30 kilograms. At the second World Humanoid Robot Games in August 2026, the gold, silver and bronze teams in weightlifting and tug-of-war heavyweight groups used Apex Hand, according to the report. Yang described six required dimensions for a dexterous hand: reliability, dexterity, payload, dynamic response, tactile sensing and precision, with reliability first. He said the hand opens and closes in about 0.2 seconds and that the damage probability for expensive motors and reducers is less than 1 percent even when a full robot falls. He said when model capability is insufficient, hardware capability is demanded; as models improve, dexterous hands can be simplified. He added that 6-DoF hands shipped more in 2023 and 2024, while high-DoF hands are now gaining attention because 6-DoF designs struggle with more complex grasping and manipulation.

Editor's Summary

On September 30, 2026, advances in robot-free data, physics-guided video generation and agent training infrastructure showed two tracks converging: better manipulation data without robot hardware, and scalable sandboxes for large agent models. Physis-Lang reported top scores on Physics-IQ Verified, while DeepSeek's DSec detailed production-scale sandbox operations. 源升智能's Apex Hand highlighted a hardware-first approach to dexterity and reliability while embodied models continue to develop.