China's two leading humanoid robotics companies have shifted their competitive battleground from manufacturing volume to open-source artificial intelligence. Just one day after Unitree Robotics open-sourced its UnifoLM-WLA-1.0 model, Shanghai-based AGIBOT fired back with Genie Envisioner Act 2.0 (GE-Act 2.0), an open-source foundation model designed specifically for robotic manipulation. This strategic pivot reveals how the race for physical AI dominance is moving beyond hardware specs and into the software layer that will ultimately determine which robots can learn and adapt fastest in real-world environments.

What's the Difference Between These New Robot AI Models?

The key distinction lies in how these models approach robotic decision-making. Most current physical AI systems, called Vision-Language-Action (VLA) models, treat robot control like text completion: they take a camera image and a language command, then directly output motor commands. This approach has a fundamental limitation: it does not explicitly model how the physical world reacts when the robot touches something.

AGIBOT's GE-Act 2.0 takes a different approach by using what researchers call a World-Action Model (WAM). Instead of jumping straight to motor commands, the system first imagines what the scene should look like after the robot acts, then works backward to figure out which movements would create that imagined future. Think of it like a chess player who visualizes the board after their move before deciding which piece to move.

The challenge with this approach is speed. Running conventional video generators inside a real-time robot control loop is notoriously slow and memory-intensive. AGIBOT solved this by breaking the task into three modular stages that work together efficiently:

• Control-First Visual Tokenizer: Compresses incoming camera frames by 64 times into a tiny grid of tokens, discarding cosmetic background details while preserving physical geometry, boundaries, and motion.
• Single-Step Visual Planner: Imagines the future visual scene in a single forward pass instead of dozens of iterative steps, generating both near-term frames for immediate control and far-horizon frames to stay on task.
• Inverse Dynamics Model: A dedicated action network translates the imagined future visual changes into smooth joint movements that the robot can execute.

How Does AGIBOT's Model Handle Ambiguous Tasks?

Physical manipulation is inherently ambiguous. If a robot is told to pick up a cup, approaching from the left or the right are both completely correct. But human demonstration datasets only record one path. This creates what AGIBOT calls the "validity gap": if the AI's visual planner imagines approaching from the right while the training demonstration approached from the left, traditional training algorithms force the robot to execute leftward movements while looking at an imagined rightward approach. This supervision clash often causes policies to freeze or average out into useless motions.

AGIBOT designed a technique called Knowledge-Aligned Selective Optimization (KASO) to solve this problem. During training, the visual planner imagines several possible futures. The action network screens them to find which candidate actually aligns with the recorded demonstration before updating the model's weights. In real-robot tests, adding KASO boosted pick success from 22.5% to 37.5%, confirming that filtering out mismatched visual futures is vital for stable learning.

What Do the Performance Numbers Actually Show?

AGIBOT tested GE-Act 2.0 across four training scales using 300, 1,200, 5,000, and 30,000 hours of interaction data. The training corpus combined teleoperation runs across both bipedal and wheeled robot platforms, synthetic simulations, open datasets, and roughly 2,000 hours of uncurated failure data. Rather than fine-tuning the model on specific test environments, which often masks whether a policy is truly generalist, AGIBOT evaluated the model out-of-the-box in zero-shot tests without any task-specific training.

The benchmark spanned 100 atomic tasks across 20 manipulation categories, featuring unseen table setups, lighting changes, and novel objects. Performance scaled steadily without flattening out. The bipedal G1-OP robot achieved success rates of 39% with 300 hours of training data, climbing to 76% with 30,000 hours. The wheeled G2-90D platform showed similar scaling, reaching 62% success with 5,000 hours and 76% with 30,000 hours.

Perhaps most impressively, the model exhibited meaningful cross-embodiment transfer: the G2-90D robot gained 17.7 percentage points across the benchmark despite making up less than 2% of the co-training dataset. Even on tasks where the G2 had under an hour of training data, such as separating paper cups or flipping coasters, the system learned the task by borrowing spatial dynamics learned on other robots.

Steps to Understanding the Scaling Implications

• Data Auditing Reveals Predictable Patterns: Zero-shot success correlated strongly with how many hours a skill category received in training. High-volume categories like wiping surfaces with 824 hours of training reached 76.7% reliability, while under-represented tasks like sweeping with a broom at 64 hours remained near 3.3%.
• Real-Time Instruction Following Works: In qualitative stress tests, when commanded to grab a green cup, the robot reached forward. Just centimeters before contact, the command was swapped to "pick up the blue cup." The arm paused for about 1.5 seconds, disengaged, and redirected to pick the blue cup, demonstrating the system can prioritize real-time human instructions over habitual routines.
• Cross-Hardware Learning Transfers Effectively: The model can learn manipulation skills from one robot platform and apply them to completely different hardware, suggesting that future robots could benefit from a shared pool of learned behaviors rather than requiring platform-specific training.

Why Does This Open-Source Battle Matter?

The release of GE-Act 2.0 directly counters Unitree's launch of UnifoLM-WLA-1.0, intensifying the race for open-source physical AI foundation models. This competition mirrors what happened in large language models, where open-source alternatives to proprietary systems accelerated innovation across the entire industry. By open-sourcing these models, both companies are betting that establishing the dominant software stack will matter more than keeping their algorithms secret.

The algorithmic battle lines in humanoid robotics are shifting rapidly from proprietary black boxes to open-source foundation models. While AGIBOT and Unitree continue to trade jabs over hardware volume, both are now racing to establish the dominant open operating stack for robotic intelligence. For the broader robotics industry, this competition could accelerate the timeline for robots that can learn new tasks quickly and transfer skills across different hardware platforms, potentially making physical AI systems more practical and affordable for manufacturers and logistics companies worldwide.