How is pricing calculated?

Pricing is based on the hourly rate of operators in your selected region. You only pay for the time operators spend controlling your robots.

What robots are supported?

We support 9+ robot arms including SO-100, Koch, Franka FR3, Franka Panda, WidowX 250, ViperX 300, ViperX 300s, and ALOHA.

What is the latency for teleoperation?

Our platform provides ultra-low latency of 30-70ms worldwide, enabling real-time responsive robot control.

ACE

ACE-Brain-0.5

ACE-Brain · July 2026

ACE-Brain Team (no institutional affiliation stated in the paper or on the arXiv page)

Reasoning VLAOpen weights, see model cardAlso written ACE-Brain-0.5-8B, ACE-Brain-0.5-VLA, ACE-Brain-0.5-Specialist

Parameters: 8.8 B; The paper says 'a single 8B backbone'.
GPU memory: 16 GB; weights at bf16, computed. Fits a 24 GB card
Inference latency: not published; per action step
Weights: see model card; downloadable checkpoint

What it is

ACE-Brain-0.5 is a unified embodied foundation model that puts spatial perception, decision making, embodied interaction, self monitoring and self improvement into one 8B backbone initialized from Qwen3-VL-8B-Instruct. It builds on ACE-Brain-0 and is trained with SSR+, which extends Scaffold, Specialize and Reconcile by a Reactivate stage after task vector merging so that grounding, navigation, manipulation and progress estimation do not interfere. For closed loop control the backbone is frozen and only a Fast Vision Pathway plus a lightweight flow matching Action Expert are trained, a configuration the paper calls ACE-Brain-0.5-VLA. Reported results include a 98.2 percent LIBERO average, 82.3 percent on SimplerEnv-Bridge and best in class progress estimation on the reversed trajectory RBM-EVAL-Refined splits. The weak spot is that breadth costs depth: on the VSI spatial video benchmark it scores 62.2 versus 63.1 for its own predecessor, and the best navigation numbers come from a separate Specialist variant rather than the unified model.

Architecture

Backbone: Qwen3-VL-8B-Instruct, via ACE-Brain-0
Action head: Flow matching Action Expert. Vision features from a Fast Vision Pathway are injected directly into the Action Expert, which transports a noise initialized action along the learned flow. For manipulation the VLM backbone is frozen and only the Fast Vision Pathway and the Action Expert are updated.
Parameters: The paper says 'a single 8B backbone'. The released checkpoint ACERobotics/ACE-Brain-0.5-8B reports 8,767,123,696 parameters in its safetensors index (Hugging Face model API, checked 11.08.2026); the model card rounds this to 9B. Stage 1 of SSR+ starts from the Qwen3-VL-8B-Instruct initialization.
Pretraining data: Not fully quantified in the paper. Named components include VSI-590K and VLM-3R for spatial and multi-view 3D understanding, AGIBOT-World (listed at 139,563 samples, 10.9 percent, 238.5M tokens, 39.2 percent), and RBM-1M for progress modeling. Robot states and actions from heterogeneous sources are canonicalized into a shared schema of joint states, gripper states and end-effector pose and rotation, with unavailable channels padded and masked.
Embodiments: Simulation only in the reported experiments: LIBERO (Franka) and SimplerEnv-Bridge (WidowX), plus VLN-CE R2R and RxR navigation in Habitat

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 8.8 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
Precision	Weights	Note
fp32	33 GB	Training master weights, rarely used for inference
bf16 or fp16	16 GB	The usual way these checkpoints are served
int8	8.2 GB	Quantised, expect some loss of precision on fine motions
int4	4.1 GB	Aggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 24 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Not stated. The paper repeatedly motivates the Action Expert by 'strict latency requirements of real-time control' but publishes no latency, no throughput, no GPU model and no memory figure. The model card documents deployment via Transformers, vLLM, SGLang and Docker without naming a device.

The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

LIBERO Average over 4 suites
98.2%
success ratePaper Table 4. Per split: Spatial 100.0, Object 100.0, Goal 96.0, Long 97.0. Produced by ACE-Brain-0.5-VLA with frozen backbone plus trained FastVision module and flow matching action expert, not the unified checkpoint.source
SimplerEnv-Bridge Average over 4 WidowX tasks, max 120-step inference horizon
82.3%
success ratePaper Table 5, ACE-Brain-0.5-VLA. Per task: Spoon 75.0, Carrot 79.2, Cube 75.0, Eggplant 100.0. Baselines are quoted from the GTA-VLA paper rather than re-run, so this is a mixed-provenance comparison.source

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

MindCube full
86.3%
accuracyPaper Table 1. Predecessor ACE-Brain-0-8B 82.1, next best embodied brain MLLM RynnBrain-8B 56.6.source
VSI full
62.2%
accuracyPaper Table 1. Beats GPT-5.4 (52.6) and Gemini 2.5 Pro (43.4) but is below its own predecessor ACE-Brain-0-8B (63.1) and well below RynnBrain-8B (71.0).source
VLN-CE R2R Val-Unseen
57.4%
success rate (SR)Paper Table 3, unified model. Navigation error 4.8 is the best among all compared general baselines. The ACE-Brain-0.5-Specialist variant reaches 62.2 SR and 4.2 NE.source
RBM-EVAL-OOD Refined (adds temporally reversed successful trajectories as negative controls)
0.88 correlation
VOC correlation rPaper Table 6. Best in table, 0.07 above Robometer-4B. The refined split causes large drops for general VLMs, for example Qwen3-VL-8B falls from 0.87 standard to 0.30 refined.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

No real robot experiments are reported. All manipulation and navigation results are in simulation. RBM-EVAL-OOD contains trajectories from unseen embodiments including Franka, but these are offline progress estimation evaluations rather than closed loop control on hardware.

Fine tuning it yourself

Not documented. The repository publishes the unified checkpoint and inference paths but no fine-tuning recipe, no data format specification and no demonstration count guidance. The manipulation configuration described in the paper freezes the backbone and trains only the Fast Vision Pathway and the flow matching Action Expert.

Where it helps, where it does not

Strengths

One checkpoint covers grounding, 3D and egocentric spatial reasoning, planning, navigation, manipulation and progress estimation, with no task specific architecture changes
Strong progress estimation, including on the RBM-EVAL-Refined splits that add temporally reversed successful trajectories as negative controls (0.80 ID, 0.88 OOD VOC)
Large margins on language grounded 3D perception: Scan2Cap 83.3 versus 14.0 for GPT-5.4 and 16.8 for Gemini 2.5 Pro
Open weights and a detailed technical report with per-benchmark tables

Limits

No institutional affiliation is stated anywhere in the paper or on the arXiv page, only 'ACE-Brain Team'
No license is declared on the Hugging Face repository or in the GitHub repository (checked 11.08.2026), which makes commercial use legally unclear
Regression against its own predecessor on VSI: 62.2 versus 63.1 for ACE-Brain-0-8B, and it trails RynnBrain-8B at 71.0
Manipulation numbers come from ACE-Brain-0.5-VLA, a configuration with frozen backbone plus separately trained action expert, not from the unified model as shipped
The best navigation results (62.2 SR, 4.2 NE on R2R Val-Unseen) come from a Specialist variant; the unified model reaches 57.4 SR
No latency, no VRAM figure, no control frequency and no action chunk size are published, despite the paper arguing for real-time control
No real robot experiments at all, so hardware transfer is unproven
SimplerEnv-Bridge baseline numbers are taken from third-party papers rather than re-run by the authors

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.

Download the client Drive a real arm now