San Francisco startup Markov Robotics has used the LTX-2.5 video generation model as the control policy for a dexterous robot hand. The robot views a scene through a camera, the model generates video of the completed task, and the controller executes the motion that video implies. The write-up appeared on 8 September 2026 in Humanoids Daily — as sponsored content, and with no published benchmarks.
Key takeaways
- LTX-2.5: 22B parameters, diffusion transformer, Gemma 4 12B text encoder
- 1:192 compression — one token covers a 32×32×8 pixel block
- Fine-tuning to predict action tokens alongside video tokens took under an hour
- Sharpa Wave hand: 22 active degrees of freedom with tactile sensing
- Community License free for companies under $10M annual revenue
Video instead of a direct policy
A typical VLA policy maps image and instruction straight to joint commands. Markov inserts an extra step — the model first generates video of what should happen, and the motion follows from that.
The logic is that a model trained on a vast corpus of footage already carries physics priors. The company says the robot matches grip force to material — glass gets gentler treatment than rock.
Training robots means teaching them how the physical world actually behaves, not just what it looks like. LTX-2.5 is the open model that gets closest to that for us.
Atharva Gundawar, CEO and co-founder of Markov Robotics.
Why LTX-2.5 specifically
The reason is arithmetic. The model packs video into fewer tokens than rivals — one token maps to a 32×32×8 pixel block, giving 1:192 compression. For closed-loop control that is the difference between affordable and impossible.
Lightricks released LTX-2.5 on 6 January 2026. The pretrained variant ships without cinematic post-training, so it stays compatible with first-person footage.
What is not known
The hard numbers are missing. There are no published manipulation success rates, action-prediction accuracy figures or closed-loop latency measurements on physical hardware. Lightricks itself writes in the model card that robotics and physical AI applications are still "developing". The open question is whether video prediction encodes real contact dynamics or only appearance.
Why it matters
If a video model genuinely carries useful priors about contact physics, the cost structure of data collection changes. Instead of hundreds of hours of demonstrations on one specific robot, a short fine-tune on a general corpus would do. That inverts the logic today's VLA policies were built on. The condition is proof nobody has presented yet, and the source material is sponsored. Without independent measurement this is a hypothesis, not a production method.
What's next
- LTX-2.5 weights and the pretrained checkpoint sit on Hugging Face in gated repositories — they require accepting the licence
- Markov Robotics has given no date for publishing manipulation benchmarks, which is the first real test of the method
- Above $10M in revenue the licence stops being free, which limits adoption among larger integrators
Sources
- Humanoids Daily — The Policy Is a Video: How Markov Robotics Runs Dexterous Manipulation on LTX-2.5
- Hugging Face — Lightricks/LTX-2.5
