← Paper Digest

Elastra: A Simulation Framework for Elastic Bodies

Youran Sun, Chugang Yi, Xingyu Ren, Haizhao Yang

November 2026

A simulated G1 on Elastra mattresses, which sink and rebound under each contact.

TL;DR Elastra simulates mattresses and trampolines in MuJoCo as grids of cells that sink and rebound under contact, checked against an independent implementation and a real mattress load curve. A residual RL policy trained in it raises a simulated Unitree G1's get-up success with HoST from 44.1% to 72.5% overall, and from 55% to 89% on a trampoline.

What they do

Robot learning environments usually treat the ground as rigid. Mattresses, cushions and trampolines instead sink where they are pressed and recover when the load moves on, so every push changes where the next contact lands. Elastra adds such surfaces to MuJoCo cheaply enough for reinforcement learning. The authors use it to adapt pretrained get-up controllers so a simulated Unitree G1 humanoid can stand up on mattresses and trampolines.

How it works

  • Surface model. Each surface is a grid of cells that move only vertically. Mattress cells sit on a spring and damper and are coupled to their neighbors by a bending term. Trampoline nodes carry load through linear-plus-cubic membrane connections to a fixed rim. The 2.0 x 1.9 m mattress needs 380 degrees of freedom, against about 10,000 for a finite-element model fine enough to follow each contact.
  • Contact. MuJoCo contact forces drive the cells, and the deformed surface sets the next contacts. A single factor scales stiffness from soft to firm.
  • Residual RL. A pretrained controller (the HoST get-up policy or the ProtoMotions tracker) stays frozen while a PPO policy adds bounded corrections to its joint targets.

Results

  • Static deflections from MuJoCo and an independent Warp implementation agree within 0.3 mm on the trampoline and 3 mm on the mattress. The fitted compression law predicts held-out loads of a real mattress with 5.9-10.8% error.
  • With HoST, success over 960 trials rises from 44.1% to 72.5%: on the mattress at stiffness factor 8 from 25% to 94%, and on the trampoline from 55% to 89%.
  • With ProtoMotions on fallen postures unseen in training, success over 400 trials rises from 19.0% to 45.0%. Gains are largest on the trampolines (10% and 15% to 52.5% and 60%).

Limitations

All robot experiments are simulated; hardware tests are left for future work. Surfaces are level and uniform, with the same stiffness and damping in every cell, unlike real supports. On the mattress at stiffness factor 4, the residual HoST policy still succeeds in only 45% of trials.

Videos

HoST alone versus HoST with the residual policy, getting up on a mattress and a trampoline.
ProtoMotions alone versus with the residual policy, on a soft mattress and a trampoline.

Figures

From Figure 1: The Unitree G1 stands up from a prone pose on a mattress simulated with Elastra, with the original HoS...
From Figure 1: The Unitree G1 stands up from a prone pose on a mattress simulated with Elastra, with the original HoST policy (top) and the residual controller adapted from it (bottom). The mattress sinks under each contact and recovers as the load moves.
From Figure 2: The two surface models. Top: a steel ball resting on the simulated mattress and trampoline. Bottom: ma...
From Figure 2: The two surface models. Top: a steel ball resting on the simulated mattress and trampoline. Bottom: mattress cells rest on local springs and dampers and are coupled to their neighbors through a bending term; trampoline nodes carry the load through membrane connections to a fixed rim.
From Figure 5: The pretrained controller keeps its weights. A residual policy, reading the robot state, the controlle...
From Figure 5: The pretrained controller keeps its weights. A residual policy, reading the robot state, the controller output and, for HoST, the surface, adds a bounded correction to its joint targets before the original PD controller.
From Figure 3: HoST get-up with 96 trials per surface. (a) Success rate of the original controller and of the residua...
From Figure 3: HoST get-up with 96 trials per surface. (a) Success rate of the original controller and of the residual policy on each surface. (b) Outcomes over all 960 trials: never reaching the standing thresholds, reaching them without holding for 1 s, or success.
Abstract

When robots interact with elastic surfaces such as mattresses and trampolines, changes in loading cause the surfaces to deform and rebound, altering subsequent contact locations and forces. We present Elastra, a simulation framework for modeling this interaction and supporting the training and evaluation of robot control policies. Elastra represents each surface using elements that move only vertically, with elastic connections and damping to model deformation and recovery. Contact forces drive surface motion, and the updated surface geometry is used to compute subsequent contact forces. We systematically validate our simulations against an independent implementation and a real mattress load curve. We further use Elastra to train the Unitree G1 humanoid to stand up on mattresses and trampolines. Using the pretrained HoST policy and ProtoMotions motion tracker as original controllers, we train residual policies through reinforcement learning to add corrections to their joint targets. Adding these residual policies increases the success rate with HoST from 44.1\% to 72.5\%, and from 55\% to 89\% on the trampoline. With ProtoMotions, the success rate increases from 19.0\% to 45.0\% for initial poses not encountered during training. The code and video are available on https://github.com/yichg/elastra.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.