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.
| Item | Details |
|---|
| Company | 1X Technologies (Norway/US) |
| Robot | NEO (Humanoid), EVE (Wheeled) |
| Parameters | 160M |
| Inference Speed | ~5Hz |
| Execution Environment | On-board GPU |
| Official Site | 1x.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
| Component | Role |
|---|
| Vision-Language Model | 160M parameters, visual-language understanding |
| Cognitive Prediction Heads | Hand position, object position prediction for improved generalization |
| Diffusion Policy Decoder | Continuous action generation |
| RL Mobility Controller | Full-body locomotion: walking, stairs, sitting/standing |
Onboard AI Stack (1X Cortex)
| Function | Description |
|---|
| LLM | Conversational control and knowledge access |
| Audio Intelligence | Speech recognition, selective listening |
| Visual Intelligence | Situation awareness, object recognition |
| Memory | Conversation 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
| Input | Processing | Output |
|---|
| Video frames | Latent representation encoding | Future frame prediction |
| Robot observations | Physical simulation | State value (success probability) |
| Action trajectories | Multiple future generation | Policy evaluation score |
Key Features
| Feature | Description |
|---|
| Action-Controllable | Controlled by precise robot trajectories, not text |
| Multiple Future Generation | Predicts various outcomes from same starting point |
| Cross-Task Transfer | Combined dataset training outperforms individual |
| Scaling | Confirmed accuracy improvement with data increase |
| Metric | Result |
|---|
| Prediction-Reality Correlation | Strong |
| Policy Selection Accuracy | 90% (when real gap is 15%+) |
| Required Accuracy | 70% 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
| Motion | Description |
|---|
| Walking | Natural gait in all directions |
| Stairs | Stereo vision-based stair climbing/descending |
| Sitting/Standing | Natural posture transitions |
| Kneeling | Support for low-height work |
| Running | Fast locomotion |
| Sidestep | Narrow 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
| Task | Description |
|---|
| Object Fetching | Search and deliver user-requested objects |
| Door Opening | Operate doors while moving |
| Tidying Up | Move objects to appropriate locations |
| Appliance Use | Operate air fryer, microwave, etc. |
| Feature | Description |
|---|
| Bracing | Support with one hand while manipulating with other |
| Bimanual Coordination | Simultaneous use of both arms |
| Full-body Utilization | Simultaneous locomotion and manipulation |
Multimodal Intelligence
| Modality | Function |
|---|
| Vision | Object recognition, scene understanding, material recognition |
| Audio | Voice commands, selective attention |
| Language | Natural language conversation, knowledge provision |
| Memory | Conversation continuity, user preference learning |
Hardware: NEO
Humanoid robot with Redwood AI installed
| Item | Spec |
|---|
| Height | 5 feet 5 inches (165cm) |
| Weight | 66 pounds (30kg) |
| Actuation | Tendon-driven (inherently safe) |
| Price | ~$20,000 |
| Release | NEO 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 |
| |
+----------------------------------------------------+
| Mode | Description |
|---|
| Autonomous | Redwood AI independently performs tasks |
| Remote Supervision | 1X experts teleoperate for complex tasks |
| Learning | Continuous model improvement from supervision data |
Training
Data Sources
| Source | Description |
|---|
| Teleoperation | Human-controlled data in homes/offices |
| Autonomous Episodes | Robot’s own execution data |
| Success/Failure Both | Learn from various outcomes |
Training Methods
| Method | Use |
|---|
| Imitation Learning | Learn basic skills from human demonstrations |
| Reinforcement Learning | Locomotion control, policy improvement |
| World Model | Fast policy evaluation and selection |
Cross-Embodiment
| Robot | Form |
|---|
| EVE | Wheeled upper-body robot |
| NEO | Bipedal humanoid |
Single Redwood model supports both platforms
Evolution
NEO Version History
| Version | Timing | Features |
|---|
| NEO Beta | 2024.08 | Initial prototype, 50-100M VLM |
| NEO Gamma | 2025.02 | Improved dexterity, Redwood AI deployment |
| NEO (Consumer) | 2025.10 | $20,000 home release |
Redwood AI Development
| Timing | Development |
|---|
| Initial | 50-100M parameter VLM |
| Current | 160M VL-Transformer + Diffusion |
| World Model | Accelerated policy evaluation with 1XWM |
Funding & Partnerships
| Item | Details |
|---|
| Total Funding | $240M+ |
| Key Investors | OpenAI, Samsung, Tiger Global |
| Strategic Partner | NVIDIA (Isaac platform) |
Comparison with Other VLAs
| Item | Redwood AI | Pi0 | GR00T N1 |
|---|
| Parameters | 160M | 3.3B | - |
| Execution Environment | On-board GPU | Server | Jetson Thor |
| Speed | 5Hz | 50Hz | 30Hz |
| Target | Consumer homes | General purpose | Humanoid |
| Price Point | $20K robot | Research | Industrial |
References
See Also