VLA GalleryAll modelsGallery

VLA Gallery / All models / New in 2026

Skild S1

Skild AI · 2026-08(Skild)/2026-09(NVIDIA)

VANew in 2026Not disclosedNot disclosed

An in-context learning VA that performs new tasks from a single video demonstration without weight updates.

View architecture diagram in the galleryOfficial page ↗

Paper

No paper · Related LocoFormer: Generalist Locomotion via Long-context Adaptation

Overview

Category
New in 2026
Type
VA
Released
2026-08(Skild)/2026-09(NVIDIA)
Organization
Skild AI
Open weights
No
Parameters
Not disclosed

Architecture

Backbone LLM
Not disclosed
Vision encoder
—
Action head
In-context learning from video demonstration prompts (Not disclosed)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
Designed from scratch as an in-context learner; takes a single video demonstration as a prompt and outputs robot actions (no weight updates); details not disclosed

Training

Pretraining accelerators
NVIDIA AI infrastructure (GPU type·count not disclosed)
Pretraining duration
Not disclosed
Pretraining steps & batch
Not disclosed
GPU/TPU-hours
Not disclosed
Fine-tuning
None (in-context learning at deployment, no weight updates)
Data type
Real-robot teleoperation + UMI + human video (egocentric) + simulation·synthetic
Data scale
Total amount not disclosed; scaling study over 1k–100k hours using subsets of the pretraining data pool
Main datasets
Not disclosed
Tasks & evaluation
Controlled study (1k–100k-hour subsets of the pretraining pool): 1k hours language 53%, ICL 43%; 100k hours language 9%, ICL 66%; trained long-horizon tasks 96%; demonstration→autonomous execution in 11 minutes (potting a plant)
Robot
Bimanual robot (model not stated)

Inference

Inference device & speed
Not disclosed (uses NVIDIA TensorRT, NVIDIA blog)
Inference memory
Not disclosed
Weights bf16 (calc.)
Not disclosed
Where it runs
Not disclosed

Reinforcement learning & base

Reinforcement learning
Not disclosed · Training method details not disclosed
Base control
Fixed · No mention of base·locomotion (appears to be bimanual manipulation, model Not stated)

Notes

Key features
An in-context learning VA that performs new tasks from a single video demonstration without weight updates. Trained on 100k h, its ICL reached 66%, ahead of a language-conditioned VLA at 9%.
Notes
Date conflict (two primary sources): Skild blog 2026-08-18, NVIDIA blog posted 2026-09-10 ('launched last week'). 50–100 hours of teleoperation to collect 380 demonstrations of long-horizon tasks (over 4 minutes). NVIDIA says one video corresponds to about 380 hands-on examples.
Verification
Confirmed

Other models in this category