Epoch RL Lab
We train agents to master environments no human has solved. Robotic hands folding origami cranes. Warehouse fleets negotiating deadlocks. Protein chains discovering fold paths in hours.
trillion interactions
open-source policies
transfer success rate
flagship agent · last run
Pretrained Policies
Select a domain — benchmark scores included
OrigamiNet-v3
Origami Crane Fold
GraspCore-xl
Multi-Object Sorting
ChainFold-v2
Protein Fold Sim
Research Output
Contact-Rich Dexterous Manipulation via Residual Physics Policies
We introduce a residual policy framework that decomposes contact-rich tasks into physics-guided primitives and learned residuals, achieving 94.2% success on origami fold benchmarks.
Origami Crane — Live Policy
Multi-Agent Deadlock Resolution at Scale: 10,000-Robot Warehouses
WareNet-v5 resolves deadlocks in warehouse fleets of up to 10,000 robots with 97.8% clearance rate, trained entirely in simulation and deployed zero-shot on real hardware.
Protein Fold Pathfinding in Hours: RL over Molecular Energy Landscapes
ChainFold-v2 discovers viable protein fold paths 40,000× faster than traditional MD simulations by treating fold discovery as a navigation problem in energy space.
Sim-to-Real Transfer Survival: A Systematic Study Across 47 Hardware Platforms
We identify the critical failure modes in sim-to-real transfer and present a domain randomization protocol that achieves 94.7% average success across 47 distinct robot platforms.
Sim-to-Real Transfer
Sim-to-Real That Survives Deployment
Most sim-to-real research dies in the lab. Epoch policies ship to production.
Adaptive Domain Randomization
Physics parameter sweeps that match real-world variance
We sample friction, mass, actuator latency, and sensor noise from distributions calibrated against 47 hardware platforms. Policies trained this way generalize without any real-world fine-tuning.
Residual Physics Integration
Learned corrections on top of analytical models
Contact-Aware State Estimation
Real-time proprioception bridging the sim-reality gap
Hardware-in-the-Loop Evaluation
Every checkpoint tested on physical hardware before release
Start fine-tuning tonight.
Install the Epoch CLI, browse 312 pretrained checkpoints, and have a fine-tuned policy running on your hardware before midnight.
# browse checkpoints
# pull a policy
# fine-tune tonight
# deploy to hardware
Python 3.9+ · PyTorch 2.0+ · CUDA 11.8+
“OrigamiNet-v3 fine-tuned on our custom gripper in 6 hours. Deployed to production the same day. No other lab ships policies that actually transfer.”
Dr. Kenji Watanabe
Robotics Lead · Preferred Networks
“WareNet-v5 cleared every deadlock in our 800-robot simulation. We were about to spend 6 months building this ourselves. Epoch saved us a year.”
Priya Krishnamurthy
CTO · FluxLogistics
“The benchmark scores aren't inflated. ChainFold-v2 hit 91.5% RMSD on our held-out protein set first try. I've never seen a pretrained policy generalize this well.”
Marcus Delacroix
PhD Candidate · MIT CSAIL