AgiBot’s robots, the A2 bipedal and G2 wheeled humanoids, are well-covered in technology media. Less covered is the platform that makes them capable: Genie Sim 3.0, AgiBot’s simulation environment for generating synthetic training data at scale. In 2025, Genie Sim 3.0 processed over 10 million synthetic manipulation scenarios, generating the training data that allows AgiBot’s robots to handle previously unseen objects, environments, and task configurations. Understanding Genie Sim explains why AgiBot claims its robots can generalize, and why simulation-based data generation is the critical infrastructure competition in Chinese robotics AI.
The Data Problem in Robot AI
Training a robot to manipulate objects requires data: demonstrations of successful grasps, placements, and manipulation sequences. Collecting this data from real robots is slow and expensive: a robot operating at 1 task per minute generates 60 demonstrations per hour. To train a model that generalizes across thousands of object types, you need millions of demonstrations, decades of robot operation time at real-world collection rates. The solution is simulation: generate synthetic demonstrations in a physics simulator at 1,000x or 10,000x real speed, using photorealistic rendering and accurate physics to create data that transfers to real robot performance. The transfer quality, how well simulation-trained policies perform on real robots, is the core technical challenge. AgiBot’s Genie Sim is its solution to this challenge.
Genie Sim 3.0 Architecture
Genie Sim 3.0 is built on NVIDIA Isaac Sim, a GPU-accelerated physics simulator with photorealistic rendering. The AgiBot additions: a procedural environment generator (automatically creating factory floors, warehouses, kitchens, and other deployment environments in randomized configurations), an object library with over 50,000 3D-scanned real objects (household items, industrial parts, packaging, tools), a randomization engine (varying lighting conditions, surface textures, object positions, and robot starting poses), and an automated data quality filter (rejecting demonstrations where the physics simulation produced physically implausible results that would corrupt training).
The 10 million scenarios processed in 2025 were not 10 million unique tasks, they were variations of a smaller set of core task types (pick and place, assembly insertion, object sorting, bin picking) with randomized parameters. The randomization is the key: a policy trained on 10,000 variations of “pick a cylinder from a bin” generalizes far better than one trained on 10,000 examples of the exact same cylinder in the exact same bin position.
Sim-to-Real Transfer
The value of simulation training depends entirely on how well the trained policy works on the real robot, sim-to-real transfer fidelity. Genie Sim 3.0’s transfer approach: domain randomization (training on such varied simulated conditions that the real world falls within the distribution), real-world fine-tuning (using a small number of real demonstrations to adapt simulation-trained policies), and adversarial validation (automatically testing policies on edge cases most likely to fail before deployment). AgiBot claims sim-to-real transfer success rates above 85% for standard manipulation tasks, meaning 85% of simulation-trained policies achieve acceptable performance on the real robot without additional real-world training. This figure, if accurate, represents state-of-the-art transfer performance for manipulation AI.
Competitive Data Arms Race
China’s humanoid companies are competing on simulation data scale as much as hardware specifications. Unitree trained UnifoLM-VLA-0 on 50,000 real trajectories, a real-world collection approach. AgiBot generated 10 million synthetic scenarios, a simulation approach. Zhiyuan Robotics announced a target of 10 million data entries for its embodied AI model (discussed separately). The fundamental question is: which data type produces better-generalizing robot AI? The answer is probably hybrid, real data for the fine details of physical interaction, simulation data for coverage of rare scenarios and scale. Companies with both real deployment data (from delivered units) and simulation infrastructure (from internal development) will have the combined dataset advantage.
Open Data Initiative
In 2025, AgiBot released a subset of Genie Sim data as open source, approximately 1 million synthetic demonstrations for common household manipulation tasks. The release positioned AgiBot as a contributor to the Chinese robotics AI commons, building goodwill with research institutions and potentially attracting talent. The open-sourced subset is training data for the tasks least commercially sensitive, household manipulation rather than industrial assembly. AgiBot retains proprietary data for the industrial scenarios that differentiate its commercial product.
AgiBot hardware: AgiBot A2 vs G2. Unitree AI comparison: Unitree UnifoLM-VLA-0.
Sources
- AgiBot: Genie Sim 3.0 announcement and 10M scenarios milestone (2025)
- NVIDIA: Isaac Sim partnership with AgiBot for simulation platform
- AgiBot: Open data release announcement (2025)
- The Robot Report: AgiBot simulation-to-real transfer methodology
