Ziran Yizhi releases Natura-1 a general-purpose robot model

Article

Ziran Yizhi releases Natura-1: After learning 1000000 hours of human actions, the robot can perform tasks it has never practiced

Beating AI news flash: embodied intelligence company Ziran Yizhi has released Natura-1, an MoE model with total parameters of 50000000000 designed to enable robots to complete operations based on instructions without repeatedly training for each new task.

Previously, robot training mainly relied on manually teleoperating real robots to collect action data, which was costly and limited in scope.Natura-1 took a different approach, first using 1000000 hours of first-person human video for pretraining, then using a small amount of robot data to adapt it to dexterous hands. This dataset contains approximately 61700000 manipulation sequences covering more than 14000 scenarios.

Ziran Yizhi first reconstructs the three-dimensional movements of fingers and the body from recordings, then annotates task objectives and manipulation steps layer by layer. The model puts language, video, and actions into the same Transformer, jointly learning how the hand should move next and how the scene will change after the action. Every step in the full manipulation sequence provides a training signal, so it does not need to extract only selected action clips. During execution, the model can also refer to previous frames and real-time feedback to adjust its movements, directly controlling dexterous hands with 20 degrees of freedom each. The team says the latency from capturing an image to outputting an action is controlled at under 100 milliseconds.

After completing general-purpose robot adaptation,Natura-1 can combine existing skills without additional training for new tasks. For example, although the training demonstrations taught it only to pull out tissues and place them on a tray, the robot can follow instructions to pull out a tissue, wipe up a water stain, and then throw it into a trash can. When asked to eliminate six colors, it can select bottles that meet the conditions. When picking up playing cards lying flat, it first pushes the card to the edge of the box and then lifts it from underneath. These demonstrations use the same model weights, with no separate fine-tuning for each task. The team also released an almost 30-minute continuous-shot video of everyday interaction.

Original link https://m.theblockbeats.info/flash/371783

Key points

Ziran Yizhi has released Natura-1, an MoE model with total parameters of 50000000000.
Natura-1 first using 1000000 hours of first-person human video for pretraining, then using a small amount of robot data to adapt it to dexterous hands.
The relevant data contains approximately 61700000 manipulation sequences, covering more than 14000 scenarios.
The team says the latency from capturing an image to outputting an action is controlled at under 100 milliseconds.
The team demonstrated operations including using tissues to wipe water stains, selecting bottles by color, and picking up playing cards, and said these tasks were not separately fine-tuned.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Natura-1 It puts language, video, and robot actions into the same model, first learning from human operation videos and then adapting with a small amount of robot data. According to the team, it can use existing skills to combine and complete new instructions.

Why it matters to readers

If the relevant performance is validated, robots may no longer need separate training for each new task, changing how they adapt to new operations.

This can be understood as follows: the robot first learns a large number of human operations, then attempts to combine the skills it has learned to complete new tasks.

Risks and unknowns

The information in the article about performance, data scale, and completing new tasks without additional training comes primarily from the team's statements.
The evidence provides no independent third-party verification of generalization capability, latency, or performance in real-world deployment.
The evidence does not state Natura-1 the commercialization, pricing, or mass-production deployment status.

Ziran Yizhi releases Natura-1 a general-purpose robot model

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center