gpt-4.1-2025-04-14· OpenAI
LMArena Human Preference 140K
what if you start off and already have opensource deepseek r1
- User
what if you start off and already have opensource deepseek r1
- Assistant
Great question! If you start with an open-source model like DeepSeek-VL-Robotics (R1), your compute requirements for building a general-purpose foundational robot can be significantly reduced—but not eliminated. Here’s a breakdown of what changes and what remains:
What DeepSeek R1 Gives You
- Pretrained Multi-Modal Model: DeepSeek R1 is already trained on large-scale vision-language-action data, giving you a strong base for perception, language understanding, and some action planning.
- Open Weights: You can fine-tune or adapt the model without training from scratch.
- Community Support: Open-source means you benefit from ongoing improvements and shared datasets.
What You Still Need to Do
1. Domain Adaptation / Fine-Tuning
- Why: Your robot’s sensors, actuators, and environment will differ from DeepSeek’s training data.
- Compute: Fine-tuning is much cheaper than pretraining. For a large model (e.g., 7B–70B parameters), expect 1,000–20,000 H100 hours depending on data size and task complexity.
2. Simulation-to-Real Transfer
- Why: Bridging the gap between simulated training and real-world deployment.
- Compute: If you use simulation (e.g., Isaac Gym, MuJoCo), you’ll need 10,000–100,000 H100 hours for robust policy learning, but this is still much less than full pretraining.
3. Reinforcement Learning / Policy Optimization
- Why: For complex tasks (manipulation, navigation), you may need to optimize policies in simulation or with real robots.
- Compute: 10,000–100,000 H100 hours depending on task complexity and sample efficiency.
4. Real-World Data Collection & Fine-Tuning
- Why: To adapt to your specific robot and environment.
- Compute: 1,000–10,000 H100 hours for incremental updates.
Total Compute Estimate (with DeepSeek R1 as Base)
Phase H100 Hours (Estimate) Domain Adaptation/Fine-tune 1,000 – 20,000 Sim2Real Transfer 10,000 – 100,000 RL/Policy Optimization 10,000 – 100,000 Real-World Fine-tuning 1,000 – 10,000 Total 22,000 – 230,000 - Low end: If your use case is close to DeepSeek R1’s training data and you don’t need much adaptation.
- High end: If you need significant adaptation, new skills, or robust sim2real transfer.
Cost Comparison
- With DeepSeek R1: 460,000 (at $2/H100 hour)
- From Scratch: 17M (see previous answer)
Key Takeaways
- Starting with DeepSeek R1 can reduce your compute needs by 10–100x compared to training a foundational model from scratch.
- Most of your compute will go into adaptation, not pretraining.
- You still need significant compute for sim2real, RL, and real-world fine-tuning, but it’s much more manageable.
If you have a specific use case or robot platform in mind, I can help you estimate even more precisely!
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.