PlasticineLab: A Soft-Body Manipulation Benchmark with Differentiable Physics

Authors: Zhiao Huang, Yuanming Hu, Tao Du, Siyuan Zhou, Hao Su, Joshua B. Tenenbaum, Chuang Gan

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate several existing reinforcement learning (RL) methods and gradient-based methods on this benchmark. Experimental results suggest that 1) RL-based approaches struggle to solve most of the tasks efficiently; 2) gradient-based approaches, by optimizing open-loop control sequences with the built-in differentiable physics engine, can rapidly find a solution within tens of iterations, but still fall short on multi-stage tasks that require long-term planning.
Researcher Affiliation Collaboration Zhiao Huang UC San Diego Yuanming Hu MIT Tao Du MIT Siyuan Zhou Peking University Hao Su UC San Diego Joshua B. Tenenbaum MIT BCS, CBMM, CSAIL Chuang Gan MIT-IBM Watson AI Lab
Pseudocode Yes def von_Mises_return_mapping(F): # F is the deformation gradient before return mapping U, sig, V = ti.svd(F) epsilon = ti.Vector([ti.log(sig[0, 0]), ti.log(sig[1, 1])]) epsilon_hat = epsilon (epsilon.sum() / 2) epsilon_hat_norm = epsilon_hat.norm() delta_gamma = epsilon_hat_norm yield_stress / (2 * mu) if delta_gamma > 0: # Yields! epsilon -= (delta_gamma / epsilon_hat_norm) * epsilon_hat sig = make_matrix_from_diag(ti.exp(epsilon)) F = U @ sig @ V.transpose() return F
Open Source Code Yes Plasticine Lab is publicly available 1. 1Project page: http://plasticinelab.csail.mit.edu
Open Datasets No The paper introduces a new benchmark/environment called Plasticine Lab, which itself serves as the data for evaluation rather than using an existing public dataset for training. The environment provides tasks and configurations.
Dataset Splits No The paper discusses an "evaluation suite" and generating "configurations" for each task. It evaluates performance on these configurations and compares different methods, which serves a similar purpose to testing. However, it does not explicitly define distinct "validation splits" or a specific validation methodology for model tuning as part of its experimental setup.
Hardware Specification Yes Table 2: Performance on an NVIDIA GTX 1080 Ti GPU. We show the average running time for a single forward or forward + backpropagation step for each scene.
Software Dependencies No The paper mentions key software components such as Taichi, Diff Taichi, MLS-MPM, and CUDA. However, it does not specify concrete version numbers for any of these software dependencies, which would be necessary for full reproducibility.
Experiment Setup Yes We train each algorithm on each configuration for 10000 episodes, with 50 environment steps per episode. We list part of the hyperparameters in Table 3 for SAC, Table 5 for TD3 and Table 4 for PPO. We fix c1 = 10, c2 = 10 and c3 = 1 for all environments reward.