Research map / Reinforcement learning / Visual and representation RL
EASI: Evolutionary Adversarial Simulator Identification for Sim-to-Real Transfer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EASI, which frames sim-to-real as a search where evolution strategies generate simulator parameter distributions and a WGAN-style discriminator over state transitions serves as the fitness function, to identify a realistic simulator from little real data.
Problem
Sim-to-real transfer suffers from the reality gap; domain randomization needs domain expertise to set ranges and yields conservative policies, while existing simulator-identification methods need much costly real data and struggle with high-dimensional control.
Approach
Combine a GAN-style adversary with evolution strategies: ES acts as the generator over simulator physical parameters, and a Wasserstein discriminator distinguishing simulated from real state transitions provides the fitness, so the parameter distribution evolves to make simulated transitions resemble real ones.
Key results
- In sim-to-sim tests, EASI improves over UDR across tasks, e.g., Ant returns ~780 vs 589 UDR under 0.25x body mass shift and Cartpole 164 vs 117 under 0.25x Cart DoF P gain.
- In real Cartpole, EASI reduces angle error to 1.67e-2 from 2.65e-2 for UDR and cart velocity error to 1.21e-1 from 1.63e-1.
- With only 1 reference trajectory, EASI still improves Ballbalance from 226.9 (UDR) to 360.1 and Go2 from 0.62 to 0.44 (table shows mixed results at low data).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| identify simulator parameters with EASI rather than the strongest listed baseline, FineTune | sim-to-sim policy transfer · Cartpole, Cart DoF P × 0.25 target | policy performance: 135.1±7.8 -> 164.4±4.6 | Helped |
| identify simulator parameters with EASI rather than the strongest listed baseline, FineTune | sim-to-sim policy transfer · Cartpole, Cart DoF P × 0.5 target | policy performance: 171.2±8.4 -> 182.9±2.2 | Helped |
| identify simulator parameters with EASI rather than the strongest listed baseline, FineTune | sim-to-sim policy transfer · Ant, Body Mass × 0.25 target | policy performance: 590.2±18.6 -> 780.7±41.5 | Helped |
| identify simulator parameters with EASI rather than the strongest listed baseline, FineTune | sim-to-sim policy transfer · Ant, Body Mass × 0.5 target | policy performance: 640.8±30.0 -> 777.4±37.5 | Helped |
| use EASI-optimized simulator parameters instead of UDR | Cartpole reality policy · Cartpole reality | Angle Error×10−2: 2.65±0.19 -> 1.67±0.16 | Helped |
| use EASI-optimized simulator parameters instead of UDR | Cartpole reality policy · Cartpole reality | Cart Vel ×10−1: 1.63±0.03 -> 1.21±0.08 | Helped |
| reduce reference data from 200 trajectories to 1 (ablation) | EASI · Cartpole target environment | policy performance: 182.9±4.3 -> 185.1±3.1 | Helped |
| reduce reference data from 200 trajectories to 50 (ablation) | EASI · Ant target environment | policy performance: 724.0±31.5 -> 747.0±25.0 | Helped |
| reduce reference data from 200 trajectories to 1 (ablation) | EASI · Ballbalance target environment | policy performance: 393.1± 60.5 -> 360.1±53.8 | No clear effect |
| vary the number of reference trajectories across 1, 50, 100, and 200 (ablation) | EASI · Go2 target environment | policy performance: 0.44±0.28; 0.80±0.06; 0.72±0.05; 0.78±0.02 | Mixed |
Surprising findings
- With 1 reference trajectory, Go2 performance under EASI (0.44) was lower than UDR (0.62), which the paper does not discuss in depth.
- GARAT performed worse than UDR in some OOD body-mass settings (e.g., Ant 228 vs 596).
Limitations
- Requires collecting task-specific state transition data per scenario and reasonable sensor noise levels; noisy sensors need imitation from imperfect demonstrations.
- Performance drops when task style differs between demonstrations and target task, so expert data is recommended per task.
- Evaluated only on a few simulated tasks and two real robots (Cartpole and Go2).
Experimental setup
Datasets: Isaac Gym simulated state transitions real-world Cartpole trajectories real-world Go2 trajectories
Benchmarks: Cartpole Go2 Ant Ballbalance real-world Cartpole real-world Go2
Metrics: episode return in target environment angle error cart velocity speed tracking L2 parameter search error joint movement frequency spectrum
Baselines: Uniform Domain Randomization Oracle FineTune GARAT
Builds on: Domain Randomization Generative Adversarial Network Evolution Strategies (mu/mu_I, lambda) WGAN GARAT
Closest papers by meaning
- DEAL: Diffusion Evolution Adversarial Learning for Sim-to-Real Transfer (NeurIPS 2025 · similarity 0.91)
- Provable Sim-to-real Transfer in Continuous Domain with Partial Observations (ICLR 2023 · similarity 0.88)
- Overcoming the Sim-to-Real Gap: Leveraging Simulation to Learn to Explore for Real-World RL (NeurIPS 2024 · similarity 0.87)
- SPiDR: A Simple Approach for Zero-Shot Safety in Sim-to-Real Transfer (NeurIPS 2025 · similarity 0.85)
- ASID: Active Exploration for System Identification in Robotic Manipulation (ICLR 2024 · similarity 0.84)
- C-GAIL: Stabilizing Generative Adversarial Imitation Learning with Control Theory (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.