# Light-O1

> Source: https://aiwiki.ai/wiki/light_o1
> Updated: 2026-09-23
> Fact-checked: 2026-09-23
> Categories: Embodied AI, Humanoid Robots, Open Source AI, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Light-O1." aiwiki.ai, 23 Sept 2026. https://aiwiki.ai/wiki/light_o1
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Light-O1** is a whole-body [foundation model](https://aiwiki.ai/wiki/foundation_model) for [humanoid robots](https://aiwiki.ai/wiki/humanoid_robot) developed by **Light Origins** (Chinese: 亮源新创), an [embodied AI](https://aiwiki.ai/wiki/embodied_ai) company with teams in Beijing, Shenzhen and Singapore. Announced on 21 September 2026,[1][12] it is pretrained on structured human actions recovered from internet video and then adapted to specific robots with purpose-collected robot data. The company calls it its "first general-purpose embodied foundation model" and reports a "cross-embodiment transfer scaling law": in its experiments, more human-action pretraining gave power-law reductions in next-action prediction loss and whole-body pose error after adaptation to a new body.[1] Alongside the technical report, Light Origins released **Light-O1-Preview**, a text-to-action checkpoint of about 5.5 billion parameters, under the [Apache License 2.0](https://aiwiki.ai/wiki/apache_license), with inference code on GitHub and a hosted playground on [Hugging Face](https://aiwiki.ai/wiki/hugging_face).[2][3]

## Overview

| Attribute | Detail |
|---|---|
| Developer | Light Origins (亮源新创) |
| Announced | 21 September 2026[1] |
| Model type | Autoregressive Transformer over interleaved language, vision and discrete action tokens[1] |
| Base model | [Qwen3.5](https://aiwiki.ai/wiki/qwen3_5)-4B[1][2] |
| Released checkpoint | Light-O1-Preview (text in, reasoning text plus whole-body action out)[2][3] |
| Parameters (Preview) | 5,510,550,016 in BF16, per the Hugging Face safetensors metadata; Pandaily rounds this to 6B[2][4] |
| Action output | 138 values per frame at 20 frames per second, 22-joint skeleton[2] |
| License | Apache 2.0 for weights and code[2][3] |
| Robots shown | LightBot (Light Origins' in-house humanoid) and [Unitree G1](https://aiwiki.ai/wiki/unitree_g1)[1] |

## Light Origins

Light Origins describes itself as a company building "foundation models for Physical AI" (see [physical AI](https://aiwiki.ai/wiki/physical_ai)).[7] Its founder and chief executive is Roger Jiang (姜旭). According to the company, Jiang worked at [OpenAI](https://aiwiki.ai/wiki/openai) on pre-training, alignment, optimization and infrastructure for [InstructGPT](https://aiwiki.ai/wiki/instructgpt), [ChatGPT](https://aiwiki.ai/wiki/chatgpt) and GPT-4, including research on [reinforcement learning from human feedback](https://aiwiki.ai/wiki/rlhf), and is credited eight times in the GPT-4 Technical Report. The company also says he holds a PhD in physics from the University of Maryland.[7] In a July 2026 interview published on the company site, Jiang said he joined OpenAI in 2019 and left in 2023.[17]

36Kr's AI outlet 智能涌现 reported in December 2024 that Jiang had founded the company in August 2024, that it was then using the English name "Light Robotics", and that it had offices in Shenzhen and Singapore. A person familiar with the matter told 36Kr that the startup was focused on consumer companion robots.[8] The company's own funding announcement says it was founded in late 2024.[6] Sina Tech identified Fan Tingxiang (范廷翔) as co-founder and chief technology officer, responsible for robot engineering and deploying algorithms on hardware.[10]

On 14 August 2026 Light Origins announced a Pre-A round of "several hundred million yuan". CAS Investment (国科投资) led the round, and China Merchants Venture Capital (招商局创投) and Xiang He Capital (襄禾资本) also took part. The company said it had more than 100 employees, over 90 percent of them in research and development, and that its hardware staff came from companies including Huawei and DJI.[6] Securities Daily reported the same round and investors and gave the company's registered name as 深圳亮源新创科技有限公司.[9]

The company organizes its research around three "scaling paradigms": scalable pre-training, scalable alignment and scalable deployment.[6][7] Light-O1 was the fourth technical release in about seven weeks, and the third in three weeks after LightNav-0 and Light REACT:[11]

| Date | Release | Description |
|---|---|---|
| 3 August 2026 | LightParkour | Parkour and locomotion skills grown in simulation from short human-motion seeds and distilled into one onboard depth policy on "Lightbot 0"[11] |
| 1 September 2026 | LightNav-0 | Open navigation model fine-tuned from Qwen3-VL-4B-Instruct and released under Apache 2.0, which the company says generalizes zero-shot across robot bodies[11][15] |
| 9 September 2026 | Light REACT | Whole-body control that uses recent physical interactions to adapt to faults, recover from falls and fall back to crawling[11] |
| 21 September 2026 | Light-O1 | Whole-body foundation model pretrained on human action, plus the open Light-O1-Preview[1] |

The company's products page says a bipedal humanoid is "coming soon" but gives no specifications.[16]

## Approach

### Why human video

The report argues that current robot action models depend on purpose-collected data. It cites more than 10,000 hours of robot demonstrations used to pretrain [π0](https://aiwiki.ai/wiki/pi0), over 500,000 hours of handheld-device capture reported by [Generalist AI](https://aiwiki.ai/wiki/generalist_ai) for GEN-1, and over one million hours of mostly egocentric human video reported by [Dyna Robotics](https://aiwiki.ai/wiki/dyna_robotics) for Dyna-2. It sets these against human video that already exists online, which it puts at more than one billion hours. The report labels that figure an "illustrative lower-bound estimate" rather than a measured inventory.[1] In public talks Jiang has put the amount of internet video at about 10 billion hours.[17]

The stated principle is "intelligence through compression". A language model uses next-token prediction to compress text into a reusable prior; Light-O1 applies the same objective to human action, meaning how people act in a situation, how their actions change the world and how they respond to feedback.[1]

### Data pipeline and action representation

The pipeline has four stages: collecting internet video of human activity, curating and segmenting it, reconstructing people's 3D motion, and adding fine-grained language annotations aligned with the motion.[1] Actions are stored in a unified humanoid representation with three decoupled parts: root trajectory, body pose and hand state.[1] In the released checkpoint this is the `human_action_138_v1` format. Each frame holds root displacement in the ground plane, pelvis height, yaw rate, 6D rotations for 22 joints and two hand-open values, in a right-handed, Y-up frame at 20 frames per second.[2]

An action tokenizer turns these continuous values into discrete tokens and back. Light Origins says it was designed to keep head direction and hand position and orientation precise.[1] The released decoder is a causal 1D-convolution finite scalar quantization (FSQ) decoder with a 65,536-entry codebook (four levels of 16).[2]

### Pretraining and adaptation

Action tokens are interleaved with language and visual observations on a shared timeline, and an [autoregressive](https://aiwiki.ai/wiki/autoregressive_model) [Transformer](https://aiwiki.ai/wiki/transformer) is pretrained on these sequences. The mixture also includes language-guided action generation, vision-language embodied reasoning and general foundation-model data.[1]

Post-training on robot data aligns the prior with a target embodiment. The report describes two execution routes:[1]

- Decoded unified human actions can drive a compatible "behavior foundation model" (BFM), the robot-side model that executes motion, such as the one Light Origins uses on LightBot.
- A diffusion-based action expert maps the prior into another robot's own action space. The report says this follows the action-expert design of [π0.5](https://aiwiki.ai/wiki/pi_0_5), a [vision-language-action model](https://aiwiki.ai/wiki/vision_language_action_model), and of GR00T N1.7.

For alignment with human intent, Light Origins says it scales reinforcement learning from human feedback. It also trains the model to answer in two parts: first a language statement of what the instruction requires of the body, then the action.[1]

## Cross-embodiment transfer scaling law

The central empirical claim concerns transfer. Starting from Qwen3.5-4B, Light Origins trained independent models at six pretraining budgets from 3.75 billion to 120 billion multimodal tokens, each budget double the last. The largest budget corresponds to 100,000 hours of human action. The base model, action tokenizer and data mixture were held fixed within the series.[1] Each checkpoint was then post-trained on three targets:[1]

- Egocentric human data from the Nymeria dataset
- Unitree G1 teleoperation data from the HIW-500 dataset
- In-house LightBot teleoperation data

For each budget the company reports the best held-out result across several post-training runs and checkpoints. Pose error is mean per-joint position error (MPJPE) under open-loop evaluation: given the prompt, earlier observations and the true action prefix, the model predicts the next one-second action chunk, scored best-of-four. Curves were fit to L(D) = L0 + a·D^(-η), following [Kaplan et al.](https://aiwiki.ai/wiki/scaling_laws), where D is the pretraining token budget and η the scaling exponent.[1]

Values published in the report's interactive figure:[1]

| Target and metric | No pretraining | 3.76B tokens | 120.3B tokens | Fitted η |
|---|---|---|---|---|
| Human action (pretraining loss, nats/token) | n/a | 6.421 | 5.672 | 0.24 |
| Egocentric human, next-action loss | 6.854 | 6.384 | 5.895 | 0.43 |
| Egocentric human, pose error (mm) | 40.11 | 39.22 | 37.78 | 0.03 |
| Unitree G1, next-action loss | 4.721 | 4.538 | 4.118 | 0.13 |
| Unitree G1, pose error (mm) | 13.34 | 13.20 | 12.75 | 0.04 |
| LightBot, next-action loss | 7.958 | 6.439 | 5.437 | 0.48 |
| LightBot, pose error (mm) | 21.39 | 19.05 | 17.61 | 0.08 |

Every pretrained checkpoint beats the no-pretraining baseline on every target. The pose-error gains are smaller than the loss gains, though. Going from no pretraining to the largest budget, pose error falls by about 4 percent on the G1, 6 percent on the egocentric data and 18 percent on LightBot, and the fitted exponents are small (0.03 to 0.08). For the egocentric and LightBot targets, pose error at 120.3B tokens (37.78 mm and 17.61 mm) is marginally higher than at 60.1B tokens (37.76 mm and 17.57 mm).[1] For the pose-error fits, the floor L0 is not fitted: it is fixed to the action tokenizer's measured reconstruction error.[1]

## Demonstrated capabilities

The report shows two kinds of behavior on real robots.[1]

**Loco-manipulation.** The instruction sets a task and the model decides how to move. Demonstrations include LightBot picking up several kinds of litter and handing them to a person, the G1 wiping water off a table while more is poured, LightBot opening a shoe cabinet and putting slippers inside after being interrupted, and the G1 following a pair of headphones as they are moved before carrying them to a desk.[1] Phoenix Tech (凤凰网科技) described LightBot as the company's self-developed small humanoid and said it could retry on its own after failures or interference.[5]

**Expressive whole-body skills.** Here the instruction names the movement itself, such as kneeling to propose with a ring, a golf swing, standing on one leg, a cartwheel, wiping a blackboard, or stepping back to pick up trash. For each example the report shows the model's reasoning trace, a rendered motion and the motion on hardware; some run on both the G1 and LightBot, others on LightBot only.[1]

## Evaluation

### Humanoid manipulation (RoboCasa GR-1)

On the simulated [RoboCasa](https://aiwiki.ai/wiki/robocasa) GR-1 tabletop benchmark (24 kitchen tasks on a bimanual humanoid with dexterous hands), all models were trained on the benchmark's teleoperated simulation data, 1,000 episodes per task, and evaluated over 50 episodes per task. The score is the macro success rate.[1]

| Model | Macro success rate | Source of result |
|---|---|---|
| Light-O1 | 79.3% | Light Origins |
| π0.5 | 70.9% | Trained by Light Origins in the OpenHLM framework |
| DIAL | 70.2% | Reported in the DIAL paper, not re-run |
| GR00T N1.7 | 59.1% | Trained by Light Origins with NVIDIA's official fine-tuning recipe |

Light Origins notes that checkpoint selection and simulator versions are not aligned between the DIAL figure and its own runs.[1] Pandaily noted that the figure comes from company-run simulations rather than an independent public leaderboard.[4] For NVIDIA's model family, see [Isaac GR00T](https://aiwiki.ai/wiki/isaac_gr00t).

### Human action generation

Light-O1 was also compared with two public text-to-motion models, HY-Motion-1.0, which the report attributes to Tencent [Hunyuan](https://aiwiki.ai/wiki/hunyuan),[1][13] and Kimodo.[14] The main benchmark was a pairwise human-rating arena over more than 30,000 prompts. An auxiliary benchmark used SSAE on HY-Motion-Bench, where a vision-language model judge checks whether the motion contains what the prompt asked for. SSAE used 1,984 prompts with four generations each.[1]

| Model | Arena Elo (range shown) | HY-Motion-Bench SSAE (overall) |
|---|---|---|
| Light-O1 | 1472.8 (1456.4-1489.4) | 78.0 |
| HY-Motion-1.0 | 1078.3 (1070.4-1086.8) | 74.7 |
| Kimodo | 1000 (anchor) | 61.4 |

Light-O1 led in all six SSAE prompt categories. Its lowest mean human rating in any motion category was above the highest category rating of either baseline, for semantic following, expressiveness and acceptability alike.[1] The report cautions that in the SSAE runs each model was run on a different date with its own default sampling settings.[1] All of these numbers are Light Origins' own measurements.

## Light-O1-Preview release

The open release has three parts: model weights on Hugging Face, inference and serving code on GitHub, and a Gradio playground on Hugging Face Spaces.[1][2][3] The Hugging Face repository was created on 20 September 2026, the day before the announcement.[2]

The Preview is narrower than the model shown in the report. Its model card calls it "a reasoning text-to-action model": it takes a natural-language prompt, writes a reasoning trace, and then emits action codes that decode to the 138-value representation. Reasoning text and actions come out in one autoregressive stream, with action codes stored as ordinary rows of the embedding table rather than handled by a separate prediction head.[2] The GitHub model table lists its context as text only.[3] The checkpoint uses a `Qwen3_5ActionForConditionalGeneration` architecture on the Qwen3.5 hybrid layout: 32 layers, with full attention in every fourth layer and linear attention in the rest.[2]

Running it locally requires Linux x86-64, Python 3.11 and an NVIDIA GPU with a CUDA 13 compatible environment. The Python API offers a [vLLM](https://aiwiki.ai/wiki/vllm) backend (the default) and a Hugging Face Transformers backend.[3] Output is an array of shape (frames, 138). It can animate a character directly or drive a behavior foundation model on a robot. The repository explains that the hardware demos used Light Origins' own BFM on LightBot and GEAR-SONIC as the motion controller on the G1. The community example included adapts the output to GEAR-SONIC for closed-loop G1 simulation in [MuJoCo](https://aiwiki.ai/wiki/mujoco); it runs in simulation only and needs a separate GEAR-SONIC policy checkpoint.[3] The model card credits Qwen3.5-4B, also Apache 2.0, as the backbone.[2]

## Reception

Pandaily called Light-O1-Preview "a downloadable motion-prior layer under a permissive license, not a turnkey robot OS". It added that the open weights let labs test the human-video pretraining claims, while "cross-embodiment loco-manipulation quality still hinges on adaptation data and controllers that remain partly proprietary."[4] Phoenix Tech's report on launch day summarized the scaling result and the LightBot household demonstrations. It also carried the company's statement that, with Light-O1, core results for all three stages of its pre-training, alignment and deployment plan were now in place.[5]

## Limitations

- The scaling law, the RoboCasa result and the motion-generation results are all self-reported. Light Origins asks users to cite the work as a company technical report hosted on its blog.[1][3]
- The published scaling series varies only the pretraining token budget, up to 120 billion tokens, on a single 4B base model. It does not measure how the trend behaves with larger models.[1]
- The released Preview generates motion from text. The listed release materials are the Preview weights, the inference code and the playground; the repository says the blog's hardware demos include Light-O1-Preview runs on LightBot and the G1, but no vision-conditioned checkpoint for the loco-manipulation tasks is listed among the release materials, and Pandaily described the adaptation data and controllers as partly proprietary.[1][2][3][4]

## References

1. Light Origins Team, "Light-O1: Scaling Whole-Body Intelligence with Human Action Pretraining", Light Origins tech blog, 21 September 2026. https://www.lightorigins.com/en/blog/light-o1
2. Light Origins, "Light-O1-Preview" model card and repository files, Hugging Face. https://huggingface.co/LightOriginsHQ/Light-O1-Preview
3. Light Origins, "Light-O1" code repository (README), GitHub. https://github.com/lightorigins/Light-O1
4. "Light Origins Open-Sources Light-O1-Preview 6B Whole-Body Model", Pandaily, 22 September 2026. https://pandaily.com/light-origins-light-o1-preview-6b-apache-human-video
5. "亮源新创发布全身智能基础模型Light-O1：让机器人从人类视频中学会全身动作", 凤凰网科技 (Phoenix Tech), 21 September 2026. https://tech.ifeng.com/c/8wbE87Kuzwg
6. Light Origins, "Light Origins Closes Pre-A Round of Several Hundred Million Yuan to Build Physical AI Foundation Models", 14 August 2026. https://www.lightorigins.com/en/news/pre-a-2026
7. Light Origins, "About". https://www.lightorigins.com/en/about
8. 田哲, "OpenAI前资深算法负责人姜旭创建「亮源新创」，打造具身智能陪伴机器人", 36Kr (智能涌现), 13 December 2024. https://www.36kr.com/p/3074151018115713
9. "亮源新创完成数亿元Pre-A轮融资 前OpenAI算法专家姜旭带队攻坚物理AGI", Securities Daily (证券日报), via Eastmoney, 14 August 2026. https://finance.eastmoney.com/a/202608143841619061.html
10. "从认路到「跌倒再战」，机器人也在重走大模型的Scaling之路？", Sina Tech (新浪科技), 12 September 2026. https://finance.sina.com.cn/tech/roll/2026-09-12/doc-inirputk6695724.shtml
11. Light Origins, "Newsroom". https://www.lightorigins.com/en/news
12. Light Origins (@LightOrigins_), announcement post on X, 21 September 2026. https://x.com/LightOrigins_/status/2101854146026717416
13. Wen et al., "HY-Motion 1.0: Scaling Flow Matching Models for Text-To-Motion Generation", arXiv:2512.23464, December 2025. https://arxiv.org/abs/2512.23464
14. Rempe et al., "Kimodo: Scaling Controllable Human Motion Generation", arXiv:2603.15546, March 2026. https://arxiv.org/abs/2603.15546
15. Light Origins, "LightNav-0" model card, Hugging Face. https://huggingface.co/LightOriginsHQ/LightNav-0
16. Light Origins, "Products". https://www.lightorigins.com/en/products
17. Light Origins, "Roger Jiang: The Key to General-Purpose Robots Resides in 10 Billion Hours of Internet Video", July 2026. https://www.lightorigins.com/en/news/dialogue-2026