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.

KST

LAPA (Latent Action Pretraining)

KAIST, South Korea, USA · October 2024

KAIST, University of Washington, Microsoft Research, NVIDIA, Allen Institute for AI

Foundation VLAOpen weights, MITAlso written LAPA, LAPA-7B-openx, Latent Action Pretraining for general Action models

Parameters: 7 B; The authors label the model 7B and name the checkpoint LAPA-7B-openx; no exact parameter count is published, so 7000 M here is...
GPU memory: 13 GB; weights at bf16, computed. Fits a 24 GB card
Inference latency: not published; per action step
Weights: MIT; MIT for both code and model weights, stated in the GitHub README and on the Hugging Face repo.

What it is

LAPA pretrains a 7B VLA on video without any robot action labels. A VQ-VAE-style quantizer learns discrete latent actions between consecutive frames, the VLM is pretrained to predict those latent tokens, and only a small final fine-tune on labeled demonstrations maps them to real robot commands. On a real Franka tabletop suite it reaches 50.1% against 43.9% for OpenVLA pretrained on the same Open-X data, and a version pretrained solely on Something-Something V2 human videos still beats Bridge-pretrained OpenVLA. Pretraining cost 272 H100-hours instead of OpenVLA's 21,500 A100-hours. The published weaknesses are concrete: worse fine-grained grasping than action-labeled pretraining, acknowledged inference latency problems with no published numbers, no parameter-efficient fine-tuning, and no inference VRAM figure anywhere. Published at ICLR 2025.

Architecture

Backbone: LWM-Chat-1M (Large World Model), a 7B vision-language model, loaded from LargeWorldModel/LWM-Chat-1M-Jax. The vision encoder stays frozen and all parameters of the underlying language model are unfrozen during latent pretraining.
Action head: Two stages. First a VQ-VAE-style action quantization model (LAQ) learns discrete latent actions between image frames; the VLA is then pretrained to predict those latent tokens autoregressively. A final supervised fine-tuning stage on a small action-labeled dataset maps latent actions to real robot actions. The default latent action generation space is 84.
Parameters: The authors label the model 7B and name the checkpoint LAPA-7B-openx; no exact parameter count is published, so 7000 M here is the authors' own rounded figure and not a measured count. The Hugging Face repo holds a 13.62 GB params file, a 1.38 GB LAQ quantizer (laq_openx.pt) and a 0.58 GB vqgan, plus a 45.7 GB pretraining jsonl that is data, not weights. The paper notes that parameter-efficient fine-tuning was left to future work, so all released fine-tuning is full fine-tuning.
Pretraining data: Three pretraining variants are reported. LAPA (Open-X), the best performing one, uses 970k Open-X trajectories with no action labels. LAPA (Bridge) uses BridgeData V2. LAPA (Human Videos) uses 200k clips from Something-Something V2, which contains 220K videos of humans manipulating everyday objects and no robot at all. Pretraining LAPA (Open-X) took 8 H100 GPUs for 34 hours, 272 H100-hours in total, at batch size 128. The authors contrast this with OpenVLA's 21,500 A100-hours and claim roughly 30 to 40 times better pretraining efficiency.
Embodiments: Franka Emika Panda 7-DoF (real-world tabletop manipulation), Language Table simulated 2-DoF pusher, SIMPLER simulated Google robot arm

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 7 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	26 GB	Training master weights, rarely used for inference
bf16 or fp16	13 GB	The usual way these checkpoints are served
int8	6.5 GB	Quantised, expect some loss of precision on fine motions
int4	3.3 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.

Published figures. These come from the model authors, not from this site.
Checkpoint download	13.62 GB

The GitHub README states that fine-tuning experiments were run with 4 A100 80 GB GPUs and that latent pretraining used 8 H100 GPUs for 34 hours, with 70K steps at batch size 256 sufficing for decent downstream performance. No inference-time GPU memory figure is published anywhere, and the authors explicitly leave parameter-efficient fine-tuning to future work, so LoRA-style low-memory adaptation is not supported out of the box.

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.

Language Table (simulation) in-domain, pretrained on 181k trajectories then fine-tuned on 1k action-labeled trajectories (0.5%), seen tasks
62%
success rateReported by the LAPA authors, Table 1 (62.0 plus or minus 8.7). VPT 44.0%, UniPi 22.0%, from scratch 15.6%. An action-labeled VLA upper bound reaches 77.0%, so LAPA closes but does not eliminate the gap to supervised pretraining here.source
Language Table (simulation) cross-environment, pretrained on 440k real-world trajectories then fine-tuned on 1k simulation trajectories, seen tasks
33.6%
success rateReported by the LAPA authors, Table 1 (33.6 plus or minus 12.7). VPT collapses to 18.0% and UniPi to 13.6%, so latent actions transfer across a real-to-sim gap better than inverse-dynamics pseudo-labeling. Error bars are wide.source

Real world

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

Real-world tabletop manipulation (Franka Emika Panda) 3 tasks, 54 rollouts per model, covering unseen object combinations, unseen objects and unseen instructions
50.1%
average success rate, LAPA pretrained on Open-XReported by the LAPA authors, Table 2. OpenVLA pretrained on Open-X and fine-tuned on the same downstream data reaches 43.9%, a third-party evaluation of OpenVLA carried out by the LAPA authors. LAPA wins on all three generalization categories: 57.8 versus 46.2 for unseen combinations, 43.9 versus 42.1 for unseen objects, 48.5 versus 43.4 for unseen instructions.source
Real-world tabletop manipulation (Franka Emika Panda) 3 tasks, 54 rollouts, Bridge-pretrained models
36.8%
average success rate, LAPA pretrained on BridgeData V2Reported by the LAPA authors, Table 2. OpenVLA (Bridge) 30.8%, an action-labeled VLA on the same data 32.6%, training from scratch 21.2%. This is the paper's central claim: label-free latent pretraining beats action-labeled pretraining on the same videos.source
Real-world tabletop manipulation (Franka Emika Panda) 3 tasks, 54 rollouts, pretrained only on Something-Something V2 human videos
34%
average success rateReported by the LAPA authors, Table 2. Pretraining on human videos with no robot and no action labels still beats OpenVLA (Bridge) at 30.8% and from-scratch training at 21.2%.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Evaluated on a real 7-DoF Franka Emika Panda tabletop setup on three tasks (knock object over, cover object with towel, pick object and put it in the sink), 54 rollouts per model, split into unseen object combinations, unseen objects and unseen instructions. LAPA (Open-X) averages 50.1%, LAPA (Bridge) 36.8% and LAPA (Human Videos) 34.0%, against OpenVLA (Open-X) 43.9%, OpenVLA (Bridge) 30.8% and from-scratch 21.2%. Fine-tuning used 450 real demonstrations across the 3 tasks.

Fine tuning it yourself

Three-stage pipeline: train the LAQ latent action quantizer, run latent pretraining on unlabeled video, then fine-tune on a small action-labeled robot dataset to map latent actions to real actions. Downstream data must be preprocessed into a JSON format containing images, instructions, raw actions and end-effector states, plus an action-scale CSV for deployment. Reference runs used 4 A100 80 GB GPUs for fine-tuning on real trajectories; the SIMPLER fine-tuning script uses only 100 trajectories and the real-world one 450. Latent pretraining used 8 H100 GPUs for 34 hours, with 70K steps at batch size 256 stated as sufficient.

Where it helps, where it does not

Strengths

Removes the action-label bottleneck: pretraining needs only video, which opens web-scale data that teleoperation cannot reach
Pretraining cost of 272 H100-hours versus OpenVLA's 21,500 A100-hours, which the authors put at roughly 30 to 40 times more efficient after adjusting for the H100 speed advantage
Beats OpenVLA on the same real Franka tabletop tasks in every generalization category when both are pretrained on Open-X (50.1% versus 43.9%)
Pretraining purely on Something-Something V2 human videos, with no robot in the data at all, still beats OpenVLA pretrained on Bridge robot data (34.0% versus 30.8%)
The authors hypothesize that avoiding ground-truth action labels prevents overfitting to a specific robot's action space, which is why LAPA adapts better across embodiments
Code and weights under MIT, with the LAQ quantizer released separately so the latent action model can be reused

Limits

Underperforms action-labeled pretraining on fine-grained motion generation such as grasping, which the authors state as their first limitation and attribute to a too-small latent action generation space (default 84)
The authors acknowledge real-time inference latency as an open problem and publish no latency, throughput or control frequency at all
No parameter-efficient fine-tuning: the paper explicitly leaves LoRA-style adaptation to future work, and the documented fine-tuning setup is 4 A100 80 GB GPUs
No inference VRAM figure is published anywhere, so hardware sizing must be measured locally against a 13.62 GB checkpoint
Only one released checkpoint (LAPA-7B-openx). The Bridge and human-video variants from the paper are not published as separate weights
On Language Table the action-labeled upper bound still wins in every column, so latent pretraining narrows but does not close the gap where labels exist
Only manipulation is explored; the authors state they have not applied the method to navigation, driving or other domains
JAX codebase built on the Large-World-Model repository, which is a different toolchain from the PyTorch mainstream around OpenVLA
Repository last pushed January 2025, with no newer release as of August 2026

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