In the λ0 team's estimate, recording people took about a quarter of the labor of robot teleoperation, cutting the data cost that limits how fast humanoid makers train robots.
Amap, Alibaba Group’s mapping and navigation unit, has released λ0 (Lambda-0), a foundation model for humanoid robots that learns walking, posture and two-handed manipulation largely from recordings of people doing everyday tasks. The model draws on HumanVerse, a 500-hour dataset of first-person video paired with body and hand motion, and completed 25 of 40 trials across four real-world tasks on a Unitree G1 humanoid, a 62.5% success rate, according to the team’s project page. TechNode reported the release on October 26, crediting Amap and Alibaba Token Hub as its developers.
The claim that λ0 learns from human data needs one qualification. In a technical report first posted on arXiv on October 13 and revised on October 21, the researchers show zero-shot transfer, meaning the robot attempted tasks for which it had seen no robot demonstrations, on seven tasks such as closing a laptop, opening a fridge and relocating objects. Those runs were scored on partial progress: 51.7% on average for whole-body tasks and 76.2% for long-range tasks. The 62.5% success figure comes from the full λ0 recipe, which adds a final training stage on 600 robot demonstrations, 150 for each task.
The cost argument is the reason the work matters. Humanoid developers usually gather training data by teleoperation, in which a person steers a physical robot through each task and then resets the scene, so the volume of data is capped by the number of robots and operators available. The report calls robot demonstrations "costly" to collect at scale. A person wearing a head camera and motion trackers can record in a home, supermarket or office with no robot present. For two representative tasks, the researchers estimate that human capture cut labor per demonstration from 140.85 to 33.30 person-seconds, about 76.4% less.
HumanVerse spans 827 task types recorded with off-the-shelf gear
HumanVerse holds 32,993 episodes and 87 million pose frames across 827 task types, including cleaning, cooking, laundry and carrying large objects, recorded in homes, supermarkets, industrial sites, offices and outdoor areas, according to the arXiv report. Collectors wore a neck-mounted GoPro HERO13 camera set at the height of the G1’s camera, a PICO 4 Ultra headset with body trackers that capture 24 body joints, and wrist trackers; hand motion was reconstructed from the video. The report does not say how many people collected the data or what the rig costs.
λ0 is a vision-language-action model, a system that takes camera images and instructions and outputs motor commands. It is built on a 2-billion-parameter Qwen3.5 backbone from Alibaba’s model family, with an action module that follows the design of the π0.5 robot model. Training runs in three stages: hand-object interaction from seven public first-person datasets, whole-body behavior from HumanVerse converted into the G1’s format, and finally robot demonstrations.
QbitAI, which on October 23 published an account supplied by Amap, put data collection at about 3.4 times the speed of mainstream methods, roughly 29 seconds per demonstration against about 100 seconds. The report itself describes that ratio as an estimate for representative tasks, "not a corpus-wide measurement."
Human recordings lifted λ0 above four rival robot models
λ0’s 62.5% success on the four real-world tasks compares with 40.0% for π0.5, 27.5% for StarVLA, 10.0% for Ψ0 and 7.5% for GR00T N1.6, according to the project page. On SIMPLE, a simulation benchmark, λ0 reached 90.0% success over 180 trials on six core tasks, 3.9 percentage points above the strongest baseline, the report says. The λ0 team ran all of these comparisons itself.
The team’s tests of each training stage show where the gains come from. Without human-data pretraining, the same setup succeeded in 15.0% of real-world trials; removing only the HumanVerse whole-body stage gave 22.5%, and removing the hand-interaction stage gave 50.0%. Task progress rose from 59.9% without whole-body training to 80.4% with the full corpus, and validation loss, a measure of prediction error, kept falling as the share of HumanVerse used grew from 5% to 100%.
Human clips carry less information each. In one test on closing a laptop and pulling a curtain, 200 human demonstrations produced a 53.3% mean task score, against 60.0% from 50 robot demonstrations. The economics therefore depend on human recordings being cheap enough to gather in far larger volumes than robot data.
Amap’s open release is still partial
TechNode reported that the project released data-collection hardware configuration code and sample data, and the project page links a HumanVerse sample on Hugging Face. Neither the report nor the project page mentions model weights or a license, and the GitHub repository linked from the report displayed "This repository is empty" when checked on October 27.
λ0 extends a robotics push that Amap began in early 2026, when it set up an embodied business unit and, on March 31, open-sourced its ABot-M0 manipulation model, Huanqiu reported. In April it showed Tutu, a quadruped guide robot, at Beijing’s Yizhuang half marathon, according to Xinhua. Alibaba’s earlier robotics exposure came mainly through investments, including in Unitree, the maker of the G1, Daily Economic News reported.
Other Chinese companies are attacking the same data constraint from different directions, including Dexforce’s bet on simulated data. The next marker is whether Amap publishes the λ0 weights and the full 500-hour HumanVerse under a stated license, which would let other G1 owners test whether recordings of people can replace robot time outside Amap’s own labs.