Amap ABot releases five models at once: embodied intelligence moves from "single-point SOTA" to "systematic competition"

Under Alibaba’s umbrella, Amap (Gaode) completely overhauled the ABot embodied system yesterday, releasing five models in one go: ABot-N1, ABot-M0.5, ABot-ER, ABot-AgentOS, and ABot-C0, claiming SOTA results on 17 authoritative benchmarks.

For developers familiar with this field, the numbers “five releases” and “17 SOTAs” aren’t surprising—benchmark-chasing in embodied intelligence has become a weekly routine. The real significance lies elsewhere: Amap didn’t just produce another stronger VLA base model this time, but rebuilt the entire system as an “organic organism.” In their own words, the five models correspond to the robot’s “hands, feet, brain, central nervous system, and motor neurons.”

That might sound like PR talk, but on closer look, it’s logically consistent.

From “stacking models” to “stacking systems”: the path of embodied intelligence has changed

Let’s start with background. In April of this year, Amap had already released a version of ABot, calling it “the world’s first full-stack embodied technology system oriented toward AGI.” It achieved 15 SOTAs, and its world model ABot-World scored above Google and NVIDIA in the CVPR 2026 Video World Model Workshop international challenge. Three months later, this upgrade feels like adding flesh to the previous skeleton—aligning everything from the underlying world model to the upper-layer Agent architecture in one go.

Over the past year, there have been two main approaches in embodied intelligence:

• One follows the “general brain” route, like π or Physical Intelligence, relying on a strong VLA foundation model to handle everything.
• The other emphasizes “body + control,” represented by companies like Boston Dynamics and Figure—pursuing physical hardware to the extreme while keeping models focused on high-level planning.

Amap takes a third path: coupling the world model, foundation model, and Agent architecture into a closed loop. The advantage of this approach is that data can self-cycle—trajectories generated in simulation are fed back to train the models, and model-derived experiences, in turn, refine the world model. In theory, the longer it runs, the faster the flywheel spins.

Of course, theory is one thing, actually running it is another. Releasing all five models together this time essentially proves that this flywheel can indeed turn.

What each of the five models does

ABot-N1: The robot’s “legs”

This is the general navigation base model. It adopts a dual-system design—a slow system for long-term logical reasoning and a fast system for real-time control. It’s reminiscent of Kahneman’s dual-system theory applied to robotics, but in practice it’s much more grounded: when navigation deviates, the model can automatically shift control from “follow the map” back to “follow the eyes.”

That design hits a real pain point. Current mainstream robot navigation either depends heavily on SLAM maps (which are brittle), or purely on end-to-end vision (which is a black box). Both can easily fail in open-world scenarios. N1’s approach enables the robot to automatically switch between “using the map when reliable” and “using vision when not.”

Another interesting innovation is its pixel-level chain-of-thought reasoning—a thought process directly overlaid on the visual feed. This isn’t just for presentation; it makes deployments traceable: you can see exactly why the robot chose a route or stopped somewhere. For embodied product teams, this is extremely valuable for debugging and A/B testing.

Metrics: 92.9% success rate in outdoor navigation, outperforming peers across five tasks—waypoint navigation, goal navigation, command following, POI navigation, and human following.

ABot-M0.5: The robot’s “hands”

A general manipulation base model. Compared with the previous version, its biggest innovation is a training method called “dream self-healing”—where the model continues to generate actions based on self-predicted “noisy visual frames.”

Let’s break this down. A major problem in traditional imitation learning is error accumulation: small step-by-step deviations compound over time during task execution, eventually causing failure. M0.5 proactively trains with “noisy visual inputs,” learning to recover correct actions from them—essentially falling into the same “pitfalls” ahead of time in its “dreams.”

As a result, the model actually performs better on complex tasks than on simple ones—exceeding previous SOTA by 20.4% on complex tasks in RoboCasa-365, and by 10.6% on basic ones. This “the harder, the better” curve is rare in embodied models, which usually exhibit the opposite.

ABot-ER + ABot-AgentOS: The robot’s “brain” and “central nervous system”

These two form the upper-layer embodied Agent architecture.

ABot-ER handles the full chain of perception-to-action reasoning—not only “what it sees,” but also “how objects, spaces, and tasks relate, and what to do accordingly.” It holds 3 SOTAs and tops the Embodied Arena 2D-EQA leaderboard.

ABot-AgentOS is more engineering-oriented. It decouples planning, invocation, execution, and verification from the robot body, connecting via plugin-style Skills. This design enables one unified system to adapt across humanoid, quadruped, and wheeled robots. For teams building robotic foundation models, the fragmentation of body types is a major challenge; AgentOS’s decoupled design offers one of the cleanest solutions so far.

Additionally, AgentOS integrates cross-task, cross-time multimodal lifelong memory and can structurally optimize its memory storage based on failed trajectories. In plain terms: when a robot trips once, it not only learns from that specific fall but can also apply that experience to similar future cases.

ABot-C0: Not detailed this time, but plays a critical role

According to official sources, C0 represents the “motor neurons,” the key module converting upper-level decisions into executable low-level control signals. Public information is limited, but judging by its name and position in the architecture, it likely bridges planning commands from AgentOS to various robot-body controllers.

Where this system adds value

Looking purely at model capabilities, 17 SOTAs is an impressive figure—but what developers truly care about is reusability.

Amap’s approach stands out in a few ways:

First, data costs have been reduced. At the base of the ABot system is ABot-World, a differentiable dynamics engine embedding physical laws into the generative process. With 3DGS technology for cold-starting, even sparse inputs like smartphone photos or aerial maps can reconstruct high-fidelity training scenes. This tackles the industry’s pain point: real-world data collection is costly, while simulation data often diverges too far from reality.

Second, semantic data is Amap’s unique advantage. With a POI database and semantic road networks accumulated from a billion monthly active users, ABot-World’s scenes are not only geometrically accurate but also semantically grounded—with “this is a Starbucks entrance” or “there’s a crosswalk ahead.” Robots learn not just paths but also why to avoid or stop somewhere.

Third, open source. Back when ABot was first released in April, Amap announced full-system open sourcing. For the embodied intelligence ecosystem, this is far more impactful than “one more SOTA.”

A measured perspective

That said, current embodied intelligence benchmarks still don’t fully reflect real-world performance. Benchmark suites like LIBERO and RoboCasa lack environmental richness, so a pretty SOTA number doesn’t necessarily translate into reliability in factories or households.

The real test for Amap’s system will be its upcoming debut in the “Tutu” robot—no matter how polished the system looks, it ultimately needs to earn real-world user approval.

From an engineering standpoint, however, ABot’s full-stack upgrade marks a leading level among domestic teams. Transitioning embodied intelligence from “single-model leaderboard play” to “systematic closed-loop design” is a direction worthy of industry-wide attention.

If you’re building embodied applications, practical points of focus might be:

• Whether the ABot-World simulation environment can serve as your training foundation
• Whether AgentOS’s Skill plugin mechanism can interface with your robot body
• Privacy and private deployment options for the lifelong memory architecture

Once the open-source code arrives, these will all have concrete answers.

In closing

By mid-2026, the embodied intelligence field has moved from last year’s “everyone can train a VLA” stage to “who can actually make a working system.” The marginal gains from one-off technical breakthroughs are shrinking, while systemic capability gaps are widening.

Amap’s ABot upgrade is the first domestic instance of a company talking about a “full-stack system” and actually delivering numbers to back it up. Whether it can ultimately fulfill its promise of real-world “seeing once, remembering forever, and inferring by analogy” will become clear once the Tutu robot hits the road.

For developers, the key signal is this: the competition dimension of embodied intelligence is shifting upward—from the model layer to the system layer. Whoever can provide a practical “embodied operating system” will have the chance to define the next-generation paradigm for robot development.