Light Origins (亮源新创)
About
About Light Origins (亮源新创)
Light Origins was founded in Singapore as Light Robotics by former OpenAI researcher Roger Jiang and now operates teams in Singapore, Shenzhen and Beijing. Its technical programme links large-scale pre-training and reinforcement-learning alignment with deployment on physical systems, so interactions from real robots can feed subsequent model development. The company is also building its own bipedal humanoid platform.
The company’s first public research release, Light-Loco-Parkour, trains a single depth-conditioned policy for walking and contact-rich traversal. Sparse human-motion seeds are repaired and expanded in simulation, combined through multi-expert distillation and transferred without real-world fine-tuning to the Lightbot 0 humanoid. The policy autonomously selects locomotion, climbing, descending and vaulting behaviours from onboard depth sensing and a velocity command.
Products
Funding
- 2026-08-14 - Pre-A - Light Origins raises several hundred million yuan in Pre-A
Light Origins completed a several-hundred-million-yuan Pre-A round led by CAS Investment for physical-AI model and robot-system development. - source: lightorigins.com
Investors
- CAS Investment - Lead - Angel / Individual - 1 funding event - first seen 2026-08-14
- China Merchants Venture Capital - Participant - VC - 1 funding event - first seen 2026-08-14
- Guoke Investment - Participant - Angel / Individual - 1 funding event - first seen 2026-08-14
- Xiang He Capital - Participant - VC - 1 funding event - first seen 2026-08-14
- Xianghe Capital - Participant - VC - 1 funding event - first seen 2026-08-14
Team
- Roger Jiang - Founder and CEO - Founder and CEO
Research
- Light Origins releases Light-O1 embodied foundation model
2026-09-21 - Research Publication - Six pretraining runs span 3.75 billion to 120 billion multimodal tokens; the largest corresponds to 100,000 hours of human action. Adaptation and evaluation use the Nymeria human-action data, Unitree G1 HIW-500 teleoperation data and in-house LightBot data. Light-O1 reports 79.3% macro success across 24 simulated RoboCasa GR-1 tasks at 50 episodes per task, and the release includes preview weights, code and a hosted playground.
Source - Light Origins publishes Light-Loco-Parkour for whole-body humanoid locomotion
2026-08-01 - Research Publication - The policy uses only onboard depth sensing and a velocity command to select walking, balancing, climbing, stepping down or vaulting without a runtime skill label, state machine or motion graph. A Real2Sim2Real pipeline expands sparse human-motion seeds across terrain variations, unifies multiple experts and distils privileged geometry into a recurrent depth policy. The same policy transferred without real-world fine-tuning to indoor and outdoor tests on Lightbot 0, including unseen obstacle variations.
Source
Timeline
Events
- 2026-09-21 - Research Publication - Light Origins releases Light-O1 embodied foundation model
Six pretraining runs span 3.75 billion to 120 billion multimodal tokens; the largest corresponds to 100,000 hours of human action. Adaptation and evaluation use the Nymeria human-action data, Unitree G1 HIW-500 teleoperation data and in-house LightBot data. Light-O1 reports 79.3% macro success across 24 simulated RoboCasa GR-1 tasks at 50 episodes per task, and the release includes preview weights, code and a hosted playground.
Source - 2026-08-14 - Funding - Light Origins raises several hundred million yuan in Pre-A
China Merchants Venture Capital, Xiang He Capital and other institutional investors participated. Capital is allocated to large-scale model training, multimodal data infrastructure, full-stack hardware and software R&D, and team expansion. The issuer disclosed only a qualitative several-hundred-million-yuan range.
Source - 2026-08-01 - Research Publication - Light Origins publishes Light-Loco-Parkour for whole-body humanoid locomotion
The policy uses only onboard depth sensing and a velocity command to select walking, balancing, climbing, stepping down or vaulting without a runtime skill label, state machine or motion graph. A Real2Sim2Real pipeline expands sparse human-motion seeds across terrain variations, unifies multiple experts and distils privileged geometry into a recurrent depth policy. The same policy transferred without real-world fine-tuning to indoor and outdoor tests on Lightbot 0, including unseen obstacle variations.
Source