IndustReal: Transferring Contact-Rich Assembly Tasks from Simulation to Reality

July 13, 2026

-solve assembly tasks in simulation with reinforcement learning
1.simulation-aware policy updates
2.signed-distance-field rewards
3.sampling-based curricula for robotic RL agents
4.policy-level action integrator to minimize error at policy deployment time

nonlinear complementarity problem (NCP)

-mathematical formulation where two quantities can't both be active at the same time
-two vectors xgeq0x\\geq 0, F(x)geq0F(x) \\geq 0 that satisfy
xTF(x)=0x^T F(x) = 0
-at least one value in every pair (xi,Fi(x))(x_i, F_i(x)) must be 00

Benchmarks:

"We solve several challenging tasks proposed in Factory [48] (pick, place, and insertion tasks for pegs-and-holes and gear assemblies in simulation) with success rates of 82-99%. We provide careful evaluations over 265k simulated trials to show the utility of SAPU, SDF-based rewards, and SBC for solving these tasks."

Setup:

-Simulation: Factory in Isaac Gym
-Robot: Franka Panda cobot
-Camera: Intel RealSense D435 RGB-D
-no F/T or task-specific grippers
-Used PPO implementation from rl-games

Rewards:

-useful for ensuring progress in long-horizon tasks
-also susceptible to reward hacking (agent gets rewards for making infinite incremental progress without reaching a terminal state)

Algorithms

1. Simulation-Aware Policy Update (SAPU)

Description

-discourages the policy from exploiting unrealistic contact behavior in simulation
-after each episode, it measures the maximum interpenetration between the plug and socket and uses this information to determine how much that episode should influence the PPO policy update.

Motivation

-physics simulators allow small collision artifacts, such as objects unrealistically passing through one another
-these may be exploited to maximize reward in simulation, but fail on real robots

Implementation

-after each rollout, compute the maximum interpenetration depth by sampling 1,000 points on the plug mesh and measuring how deeply they penetrate the socket mesh
-during the PPO update, trajectories with excessive penetration are filtered out, downweighted, or both

2. Signed Distance Field (SDF)-based dense reward

-instead of using keypoints for dense rewards, compare the entire geometry
-SDF - positive outside object, 0 on surface, and negative inside object
-before training, sample 1000 points on the receptacle surface
-place the object in the target pose, and precompute the distance to the receptacle using an SDF with the 1000 points
-compute RMS of the SDF

sampling-based curriculum (SBC)

-prevent overfitting to early stages of a curriculum

Policy-Level Action Integrator (PLAI)

-closed-loop integration of policy actions to compensate for real-world execution error
-instead of applying actions to the current state, actions are accumulated on the previous desired state
st+1d=stdoplusat=stdoplusPi(ot)s^d_{t+1} = s^d_t \\oplus a_t = s^d_t \\oplus \\Pi(o_t)
-acts similarly to the integral term of a PID controller, reducing steady-state error caused by friction, gravity mismatch, and other unmodeled dynamics

Results

-~87-100% for partial insertions across pegs, gears, and connectors
-lower accuracy on full insertions