VLA Gallery / All models / Industry models
Skild Brain
Skild AI · 2025-07
기타Industry modelsNot disclosedNot disclosed
Positions itself as an omni-bodied model controlling quadrupeds·humanoids·tabletop arms·mobile manipulators in one.
Paper
No paper
Overview
- Category
- Industry models
- Type
- 기타
- Released
- 2025-07
- Organization
- Skild AI
- Open weights
- No
- Parameters
- Not disclosed
Architecture
- Backbone LLM
- Not disclosed
- Vision encoder
- —
- Action head
- Low-frequency high-level → high-frequency low-level policy (Not disclosed)
- Chunk
- Not disclosed
- Control rate
- Not disclosed
- Architecture (original summary)
- Two tiers: low-frequency high-level manipulation · navigation action policy → high-frequency low-level action policy; backbone and action representation not disclosed
Training
- Pretraining accelerators
- Not disclosed
- Pretraining duration
- Not disclosed
- Pretraining steps & batch
- Not disclosed
- GPU/TPU-hours
- Not disclosed
- Fine-tuning
- Not disclosed
- Data type
- Simulation·synthetic + human video (internet video) + real robot (post-training)
- Data scale
- Not disclosed (blog only describes the required scale as 'trillions of examples')
- Main datasets
- Not disclosed
- Tasks & evaluation
- Demos: climbing stairs, precise part assembly, moving on slippery slopes, household chores such as dishwashers; quantitative benchmarks not disclosed
- Robot
- Quadruped, humanoid, tabletop arm, mobile manipulator
Inference
- Inference device & speed
- Not disclosed (two-tier low-frequency·high-frequency policy)
- Inference memory
- Not disclosed
- Weights bf16 (calc.)
- Not disclosed
- Where it runs
- Not disclosed
Reinforcement learning & base
- Reinforcement learning
- Not disclosed · Simulation·internet video pretraining. RL use not disclosed
- Base control
- Quadruped·Humanoid legged (Commands/latents → RL locomotion controller) · High-level manipulation·navigation policy → low-level high-frequency policy outputs joint angles·motor torques (quadruped·humanoid locomotion; mobile manipulator base format not stated)
Notes
- Key features
- Positions itself as an omni-bodied model controlling quadrupeds·humanoids·tabletop arms·mobile manipulators in one. The company stresses that the model is not 'a VLM in disguise'.
- Notes
- Type is Other since VA/VLA status is not disclosed. General-purpose model including locomotion, so only the manipulation part is compared. Announced 2025-07-29 (blog·press release). Blog has no statements on parameters·data hours·compute.
- Verification
- Confirmed
Other models in this category
- Gemini Robotics2025-03
- Gemini Robotics-ER2025-03
- Gemini Robotics On-Device2025-06
- Gemini Robotics 1.52025-09
- Gemini Robotics-ER 1.52025-09
- RT-H2024-03
- AutoRT2024-01
- SARA-RT2023-12
- GR-22024-10
- GR-32025-07
- GR-RL2025-12
- GR-Dexter2025-12
- TRI LBM 1.02025-07
- GO-12025-03
- Genie Envisioner2025-08
- RFM-12024-03
- DYNA-12025-06
- GEN-02025-11
- Galaxea G02025-08
- GigaBrain-02025-10
- RynnVLA-0012025-08
- UnifoLM-WMA-02025-09
- V-JEPA 2-AC2025-06