Redwood AI (1X Technologies)

1X Technologies' Vision-Language-Action Model for NEO Humanoid

Redwood AI (1X Technologies)

Home > Models > Redwood AI


Key Significance

  • First Consumer Humanoid Deployment: VLA deployed in actual homes with NEO robot - ~$20,000 price point
  • On-board Execution: 160M parameters running at 5Hz on robot’s built-in GPU, no cloud dependency
  • World Model Innovation: 1XWM predicts task success rate before actual execution, accelerates policy selection
  • Cross-Embodiment: Single model supports both EVE (wheeled) and NEO (humanoid)
  • Hybrid Operation: AI autonomous + remote expert supervision hybrid architecture
  • Safety-First Design: Tendon-driven mechanism for inherent safety
  • OpenAI Investment: $240M funding with OpenAI as major investor

Overview

Redwood AI is a VLA model developed by 1X Technologies for humanoid robots. A 160M parameter vision-language transformer running on NEO robot’s built-in GPU, performing various tasks in home environments.

ItemDetails
Company1X Technologies (Norway/US)
RobotNEO (Humanoid), EVE (Wheeled)
Parameters160M
Inference Speed~5Hz
Execution EnvironmentOn-board GPU
Official Site1x.tech/ai

Architecture

Redwood AI has a VLM + Diffusion Policy + RL Controller structure.

+-------------------------------------------------------------+
|                   Redwood AI Architecture                    |
+-------------------------------------------------------------+
|                                                              |
|   +----------+  +----------+  +----------+                   |
|   |  Vision  |  |  Audio   |  | Language |                   |
|   |  Input   |  |  Input   |  |  Input   |                   |
|   +----+-----+  +----+-----+  +----+-----+                   |
|        |             |             |                          |
|        +-------------+-------------+                          |
|                      |                                        |
|        +-------------v-------------+                          |
|        |   Vision-Language Model   |  160M parameters         |
|        |   (Cognitive Prediction)  |                          |
|        +-------------+-------------+                          |
|                      |                                        |
|        +-------------v-------------+                          |
|        |    Diffusion Policy       |  Action decoding         |
|        |    Decoder                |                          |
|        +-------------+-------------+                          |
|                      |                                        |
|        +-------------v-------------+                          |
|        |    RL Mobility            |  Full-body control       |
|        |    Controller             |                          |
|        +-------------+-------------+                          |
|                      |                                        |
|                      v                                        |
|               NEO Robot Actions                               |
|                                                              |
+-------------------------------------------------------------+

Model Components

ComponentRole
Vision-Language Model160M parameters, visual-language understanding
Cognitive Prediction HeadsHand position, object position prediction for improved generalization
Diffusion Policy DecoderContinuous action generation
RL Mobility ControllerFull-body locomotion: walking, stairs, sitting/standing

Onboard AI Stack (1X Cortex)

FunctionDescription
LLMConversational control and knowledge access
Audio IntelligenceSpeech recognition, selective listening
Visual IntelligenceSituation awareness, object recognition
MemoryConversation continuity, past context recall

1X World Model (1XWM)

Core innovation of Redwood AI - Physics-based generative simulator

Concept

Predict task success rate without actual execution
    |
Quickly compare policy candidates
    |
Select optimal checkpoint

Architecture

InputProcessingOutput
Video framesLatent representation encodingFuture frame prediction
Robot observationsPhysical simulationState value (success probability)
Action trajectoriesMultiple future generationPolicy evaluation score

Key Features

FeatureDescription
Action-ControllableControlled by precise robot trajectories, not text
Multiple Future GenerationPredicts various outcomes from same starting point
Cross-Task TransferCombined dataset training outperforms individual
ScalingConfirmed accuracy improvement with data increase

Performance

MetricResult
Prediction-Reality CorrelationStrong
Policy Selection Accuracy90% (when real gap is 15%+)
Required Accuracy70% sufficient for valid policy selection

Limitations

  • Accuracy drops on objects not in training data
  • Difficulty with locomotion tasks having cumulative position errors

Mobility Controller

RL-based controller for NEO’s full-body locomotion

Supported Motions

MotionDescription
WalkingNatural gait in all directions
StairsStereo vision-based stair climbing/descending
Sitting/StandingNatural posture transitions
KneelingSupport for low-height work
RunningFast locomotion
SidestepNarrow space navigation

Training Method

Motion Capture data -> Kinematic Planner -> Human-like trajectory generation
                              |
               RL Controller -> Balance maintenance while following trajectory
  • Fully Simulation Trained: Real-world robustness via physics randomization
  • 2-Stage Design: High-level motion planning + Low-level balance control

Capabilities

End-to-End Mobile Manipulation

TaskDescription
Object FetchingSearch and deliver user-requested objects
Door OpeningOperate doors while moving
Tidying UpMove objects to appropriate locations
Appliance UseOperate air fryer, microwave, etc.

Multi-Contact Manipulation

FeatureDescription
BracingSupport with one hand while manipulating with other
Bimanual CoordinationSimultaneous use of both arms
Full-body UtilizationSimultaneous locomotion and manipulation

Multimodal Intelligence

ModalityFunction
VisionObject recognition, scene understanding, material recognition
AudioVoice commands, selective attention
LanguageNatural language conversation, knowledge provision
MemoryConversation continuity, user preference learning

Hardware: NEO

Humanoid robot with Redwood AI installed

ItemSpec
Height5 feet 5 inches (165cm)
Weight66 pounds (30kg)
ActuationTendon-driven (inherently safe)
Price~$20,000
ReleaseNEO Beta (2024.08), NEO Gamma (2025.02)

Safety

  • Tendon-Driven: Motors separated from joints for flexibility on collision
  • Lightweight Design: 30kg for safe human interaction
  • Compliance: Control that yields to external forces

Hybrid Operation

1X’s unique approach - AI autonomous + human supervision

+----------------------------------------------------+
|                  Hybrid Operation                   |
+----------------------------------------------------+
|                                                     |
|   NEO Autonomous Execution                          |
|        |                                            |
|        +-- Success -> Task complete                 |
|        |                                            |
|        +-- Difficulty -> 1X Expert remote supervision|
|                      |                              |
|                      +-- New skill learning         |
|                                                     |
+----------------------------------------------------+
ModeDescription
AutonomousRedwood AI independently performs tasks
Remote Supervision1X experts teleoperate for complex tasks
LearningContinuous model improvement from supervision data

Training

Data Sources

SourceDescription
TeleoperationHuman-controlled data in homes/offices
Autonomous EpisodesRobot’s own execution data
Success/Failure BothLearn from various outcomes

Training Methods

MethodUse
Imitation LearningLearn basic skills from human demonstrations
Reinforcement LearningLocomotion control, policy improvement
World ModelFast policy evaluation and selection

Cross-Embodiment

RobotForm
EVEWheeled upper-body robot
NEOBipedal humanoid

Single Redwood model supports both platforms


Evolution

NEO Version History

VersionTimingFeatures
NEO Beta2024.08Initial prototype, 50-100M VLM
NEO Gamma2025.02Improved dexterity, Redwood AI deployment
NEO (Consumer)2025.10$20,000 home release

Redwood AI Development

TimingDevelopment
Initial50-100M parameter VLM
Current160M VL-Transformer + Diffusion
World ModelAccelerated policy evaluation with 1XWM

Funding & Partnerships

ItemDetails
Total Funding$240M+
Key InvestorsOpenAI, Samsung, Tiger Global
Strategic PartnerNVIDIA (Isaac platform)

Comparison with Other VLAs

ItemRedwood AIPi0GR00T N1
Parameters160M3.3B-
Execution EnvironmentOn-board GPUServerJetson Thor
Speed5Hz50Hz30Hz
TargetConsumer homesGeneral purposeHumanoid
Price Point$20K robotResearchIndustrial

References


See Also