Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World¶
Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., & Abbeel, P. (2017). Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World. IEEE, 23-30.
Cited by¶
3 citations across 3 artifacts.
Each citation links to the sentence it supports in the citing article.
Primes¶
- Idealized-Substrate Fallacy
- The same structure governs reinforcement-learning sim-to-real transfer: the simulator is the idealized substrate (perfect actuation, no sensor noise, no latency), the robot hardware is the real substrate, and a policy that flies in sim stalls on the elided actuation-noise and latency terms — fixed by domain randomisation and latency modelling, which is the substrate audit re-staging each dropped term.
This sourceAddresses the sim-to-real gap — simulators elide actuation noise, sensor noise, and latency — by re-introducing those terms via domain randomization.
Supported in partVerified against the work's full text
Tobin et al. say simulators miss real physical effects and sensor noise, causing the reality gap, but do not discuss latency, perfect actuation or adversaries.
“the real world has unmodeled physical effects like nonrigidity, gear backlash, wear-and-tear”
- The same structure governs reinforcement-learning sim-to-real transfer: the simulator is the idealized substrate (perfect actuation, no sensor noise, no latency), the robot hardware is the real substrate, and a policy that flies in sim stalls on the elided actuation-noise and latency terms — fixed by domain randomisation and latency modelling, which is the substrate audit re-staging each dropped term.
- Training Serving Skew
- Machine-learning operations (canonical): a model trained on one period's data deployed against a later, shifted distribution; or training- and serving-time feature pipelines that process inputs differently, so the model receives different inputs at serving than at training. Sim-to-real robotics: a controller trained in a physics simulator whose missing details — friction, sensor noise, latency — cause systematic failures on the physical robot.
This sourceIntroduces domain randomization as a remedy for the sim-to-real gap in robot control.
Supported in partVerified against the work's full text
Tobin et al. support the sim-to-real half: simulator/real-world discrepancies, including unmodeled effects and noiseless sensors, hinder transfer; the ML-operations half is not in this paper.
“Unfortunately, discrepancies between physics simulators and the real world make transferring behaviors from sim- ulation challenging.”
- Machine-learning operations (canonical): a model trained on one period's data deployed against a later, shifted distribution; or training- and serving-time feature pipelines that process inputs differently, so the model receives different inputs at serving than at training. Sim-to-real robotics: a controller trained in a physics simulator whose missing details — friction, sensor noise, latency — cause systematic failures on the physical robot.
Mechanisms¶
- Digital Twin Trial
- The twin is a model, and its central hazard is the sim-to-real gap: the policy passes on the replica and fails in the world because the replica omitted the very dynamic that mattered.
This sourceIdentifies the reality gap separating simulation-trained systems from their performance on physical hardware.
- The twin is a model, and its central hazard is the sim-to-real gap: the policy passes on the replica and fails in the world because the replica omitted the very dynamic that mattered.
Verification¶
Does it exist? Confirmed. This work's DOI resolves to a registered record, which fixes its identity. That is all it fixes.
Does it back the claim? Read against the text for 2 of 3 citations: 2 supported in part. Each verdict is shown under its citation below, with what in the work backs the sentence.
Support is checked per citation rather than per work — the same source can be cited soundly in one article and wrongly in another. Per-citation recording began recently, so a citation with no recorded check is a gap in the record rather than evidence it went unchecked.
See how references were verified.
Links previously used in the corpus¶
Before the registry existed this work was also linked 1 other way.
Registry ID ref:23089ef1f4e3 · see in the full table