CARI4D vs ABot-Earth 0.7: Two 4D Models That Never Compete

If you searched for CARI4D against ABot-Earth 0.7 expecting a head-to-head, the honest answer is that no such comparison exists and none is coming. These two systems do not do the same job, do not take the same input, do not output the same thing, and would never be benchmarked against each other by anyone who understood either one. CARI4D is NVIDIA's category-agnostic 4D reconstruction of human-object interaction — it watches one person handling one object in an ordinary video and recovers their metric-scale poses over time. ABot-Earth 0.7 is Amap's 3D-native urban world model — it takes a satellite image or a text prompt and generates a kilometre-scale, explorable city as a Gaussian-splatting scene. The reason this page exists anyway is that the axis separating them is a genuinely useful one, and the two are far more different in what you can do with them than in what they appear to be.

There is also a second, sharper reason. These two releases are at opposite ends of a spectrum that matters more than any spec line: one of them you can download and run this afternoon, and the other you cannot run at all.

The answer nobody searching this query wants

Both are described as 4D. Both are from major vendors. Both were released inside the last year. That is roughly where the overlap stops.

CARI4D reconstructs. Given video, it estimates what was there. The paper — Category Agnostic 4D Reconstruction of Human-Object Interaction by Xianghui Xie, Bowen Wen, Yan Chang, Hesam Rabeti, Jiefeng Li, Ye Yuan, Gerard Pons-Moll and Stan Birchfield — appeared on arXiv as 2512.11988 in December 2025 and was accepted at CVPR 2026. Its subject is a human and a rigid object in contact, and its core claim is scale: recovering metric dimensions from a monocular camera with no depth sensor and no multi-view rig, which is the part prior work struggled with.

ABot-Earth 0.7 generates. Given a georeferenced satellite image or a text description, it produces a scene that did not exist as data before. Amap, the Chinese mapping and navigation company, unveiled it on September 10, 2026 at its Street Stars gala in Hangzhou, succeeding ABot-Earth 0.5 from June 8. It outputs 3D Gaussian splats, exports to Unreal Engine and Unity, and covers what Amap says is 196-plus countries and regions, up from 190-plus in 0.5.

One is an estimator. One is a generator. That distinction does more work than any benchmark table could.

A person and a chair, or a city

The scale gap is the one people underestimate. CARI4D's unit of interest is a human body and a hand-held object, measured in centimetres, tracked across the frames of a single clip. ABot-Earth 0.7's unit of interest is a square kilometre of city, generated in roughly ten minutes on a single consumer-grade GPU by Amap's own account.

• Input — CARI4D a monocular RGB video, MP4 vs ABot-Earth 0.7 a satellite image or a text prompt

• Output — CARI4D per-frame human pose and shape, object rotation and translation, and two hand-contact logits vs ABot-Earth 0.7 a 3D Gaussian splatting scene, exportable to Unreal Engine and Unity

• Unit of analysis — CARI4D one human and one rigid object in contact vs ABot-Earth 0.7 terrain, buildings, vegetation and landmarks at city scale

• Temporal claim — CARI4D reconstructs observed motion frame by frame vs ABot-Earth 0.7 is positioned as a world model with a prediction layer for what happens next

• Metric scale — CARI4D metric scale recovered from monocular video via UniDepth and a coarse-to-fine scale search vs ABot-Earth 0.7 georeferenced from a satellite image, no metric recovery problem

• Run cost — CARI4D PyTorch on an NVIDIA Ampere GPU, validated on A100, 38 hours on 8×A100 for the research training run vs ABot-Earth 0.7 about ten minutes per km² on one consumer-grade GPU, per Amap's own comparison

Notice that neither column is better than the other on a single row. They are measuring different worlds. A model that recovers the contact point between a hand and a gas cylinder is not improved by also knowing where the buildings are, and a model that generates a plausible skyline is not improved by knowing the exact wrist angle of a pedestrian.

The asymmetry that actually decides it

Here is the difference that a buyer will feel within the first hour, and it has nothing to do with capability.

CARI4D has code and weights you can obtain today. The NVlabs repository carries the official implementation, released February 28, 2026, with training code and custom-video preprocessing added April 4. The pretrained CoCoNet checkpoint downloads from Hugging Face, the Docker image xiexh20/cari4d is published, and since August 30, 2026 a commercially licensed 231M checkpoint has been available as nvidia/cari4d_commercial under the NVIDIA Open Model Agreement — gated, but obtainable. The dependencies are documented, including the awkward parts: NLF torchscript weights, FoundationPose weights, SMPL-H assets, a kid_template.npy from AGORA, and Hunyuan3D for the object meshes. You can run inference on the BEHAVE demo, on a provided in-the-wild clip, or on your own video, and evaluate it with the shipped scripts.

ABot-Earth 0.7 is not obtainable in any of those senses. Amap has published no model weights, no model card, no public API, no pricing and no developer documentation. The experience site is live but Amap describes generation as invite-only or whitelisted, the interface is Simplified Chinese only, and the 0.5-era GitHub repository reportedly held a technical report and nothing executable. Every headline number — ten minutes per square kilometre, roughly 1,000× the efficiency of traditional reconstruction, about one hundredth of the cost, 196-plus countries — traces to Amap and has not been independently reproduced by anyone outside Amap. Those are vendor-reported figures, and there is currently no route by which they could become anything else.

So the practical comparison is not "which model is better." It is that one of these is a research artefact with a commercial licence and a Docker image, and the other is a product demonstration with a press cycle. Both are legitimate achievements. Only one is something you can put in a build.

Where the two would genuinely meet

There is a real composition here, and it is worth being concrete about it, because it is the only sense in which these belong in the same conversation.

CARI4D's output is human-object interaction in a metric world frame. ABot-Earth 0.7's output is an environment. Populate the second with the first and you have something neither produces alone: a generated city in which people are doing things that are physically measured rather than artistically placed. Robotics simulation, retail layout planning and interaction-dataset generation all want exactly that combination — an environment cheap enough to regenerate and human motion accurate enough to train against.

The seam between them is a coordinate transform and a scale convention, and it is not trivial, because CARI4D's metric frame is defined relative to a single camera and ABot-Earth 0.7's is defined by geolocation. Nobody has published that integration. It is the kind of thing a team builds in a week and does not write up.

The step that gets forgotten in that pipeline is the part that turns raw interaction data into something queryable — captioning clips, tagging contact events, building retrieval indices and quality-control filters over the output. That work runs through vision-language and language models, and it is the layer OrcaRouter covers: more than 200 models behind one API, the provider's list price passed through at 0% markup, automatic failover when a provider degrades, and a routing DSL that composes several models into a single call so the captioning and the QC pass do not have to be separate integrations. Neither CARI4D nor ABot-Earth 0.7 is served there, and neither is an LLM — but the tooling you surround them with almost certainly is.

Which one you actually want

Pick CARI4D if you have video of a person interacting with an object and you need their poses, in metric units, frame by frame — for motion analysis, robotics perception, animation reference, or building an interaction dataset. It is available now, it is documented, and the commercial licence landed on August 30, 2026. Budget for the dependency chain, and read the licence terms on the parts NVIDIA does not own.

Pick ABot-Earth 0.7 if you need environments rather than actors — city-scale scenes generated from a location or a description, quickly, exportable into a game engine. Understand that you are evaluating a demonstrated capability rather than adopting a tool, that access is discretionary, and that the efficiency figures are Amap's own.

Pick both only if you are building the composed system described above, and go in knowing you will be writing the seam yourself.

The question worth asking is not which of these two wins. It is whether the thing you are actually building needs a reconstruction model, a generation model, or neither — and on that question these two answers never overlap.