Embodied Robot Data Tradeoff

Definition

Embodied robot data tradeoff is the design choice between simulation, real robot data, human first-person data, internet video, and task-specific post-training data when training robots for physical action.

Current Synthesis

The four-way Shizilukou Crossing episode makes the data question less binary than “simulation versus real machines.” 速度科技 / Sudu Technology argues that simulation can scale manipulation pretraining and reinforcement learning when physical consistency is good enough. 蚂蚁灵波 / Ant Lingbo emphasizes that real sensors, tactile signals, contact noise, and physical imperfection still contain information that simulation cannot simply invent. Poke Robotics adds a warning that raw million-hour data goals may be overvalued if they do not produce stronger general robot capability.

Key Claims

• Simulation is strongest when it cheaply expands object, environment, lighting, and task combinations for pretraining basic manipulation.
• Real robot data remains necessary because sensors, tactile readings, latency, contact, and material imperfections differ from idealized environments.
• The main dispute is not whether simulation or real data matters, but when each data type is useful in the training and deployment cycle.
• Data scale is not automatically capability scale; data quality, task diversity, feedback, and model paradigm shape whether extra hours help.
• Commercial scenes need data routes that lower post-training and deployment cost, not only benchmark performance.

Evidence

• Simulation evidence: 当具身智能走到十字路口｜对谈苏度、蚂蚁灵波、自变量、破壳：四种一线判断 records Han Zheng saying Sudu rebuilt large-scale data and simulation pipelines and uses object, environment, and lighting combinations to approach high success on grasping, placing, and assembly tasks.
• Real-data evidence: 当具身智能走到十字路口｜对谈苏度、蚂蚁灵波、自变量、破壳：四种一线判断 records Shen Yujun saying real sensors include frequency, amplitude, consistency, and physical-world imperfections that simulation struggles to match.
• Stage evidence: 当具身智能走到十字路口｜对谈苏度、蚂蚁灵波、自变量、破壳：四种一线判断 has Shen distinguish pretraining data, real-world collection, and task-specific simulation use rather than choosing one source.
• Scale qualification evidence: 当具身智能走到十字路口｜对谈苏度、蚂蚁灵波、自变量、破壳：四种一线判断 records Xu Huazhe saying very large robot-hour claims have not obviously outperformed smaller, better-structured data sets.

Counterevidence & Qualifications

The source does not provide benchmark tables, simulator transfer rates, or audited data-hour comparisons. Each participant speaks from a company route, so the synthesis should preserve strategic disagreement rather than collapse it into one data recipe.

What Changed

• Added a dedicated concept for the simulation-real-data tradeoff exposed by the four-company debate.
• The current judgment treats data type, data quality, task stage, and deployment economics as jointly decisive.

Related Concepts

• Robot Data Scale Up - broader question of reaching a scalable robot-data path.
• Sim2Real - transfer route where simulation must survive physical constraints.
• Real Robot Data Strategy - real-machine branch qualified by simulation and data-quality limits.
• Embodied Data Pyramid - adjacent recipe view for combining multiple data sources.
• Robotics Simulation Evaluation - infrastructure layer that can test and train but still needs real-world grounding.