Model Tensor Planning

Authors: An Thai Le, Khai Nguyen, Minh Nhat VU, Joao Carvalho, Jan Peters

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through experiments on various challenging robotic tasks, ranging from dexterous in-hand manipulation to humanoid locomotion, we demonstrate that MTP outperforms standard MPC and evolutionary strategy baselines in task success and control robustness. Design and sensitivity ablations confirm the effectiveness of MTP s tensor sampling structure, spline interpolation choices, and mixing strategy.
Researcher Affiliation Collaboration 1Intelligent Autonomous Systems, Department of Computer Science, Technical University of Darmstadt, Germany 2Systems AI for Robot Learning, German Research Center for AI (DFKI), Germany 3Hessian Center for Artificial Intelligence (hessian.AI), Germany 4Centre for Cognitive Science, Technical University of Darmstadt, Germany 5Automation & Control Institute, TU Wien, Austria 6Austrian Institute of Technology (AIT), Vienna, Austria 7Vin Robotics and Vin University, Vietnam
Pseudocode Yes Algorithm 1: Sampling Paths From G(M, N) ... Algorithm 2: Model Tensor Planning
Open Source Code No The paper discusses its implementation details like being fully vectorized using JAX and compatible with Mu Jo Co XLA, and mentions that 'Experiment videos are publicly available at https://sites.google.com/view/tensor-sampling/', but it does not provide concrete access to the source code for the methodology described in the paper.
Open Datasets Yes All algorithms and environments are implemented in Mu Jo Co XLA (Todorov et al., 2012; Kurtz, 2024)... Comparison Environments. Push T (Chi et al., 2023), Cube-In-Hand (Andrychowicz et al., 2020) ... Walker (Towers et al., 2024)... All tasks are implemented in hydrax (Kurtz, 2024).
Dataset Splits No The paper describes control tasks within robotic environments and specifies that experiments are run 'over 5 seeds' or '4 random seeds' for statistical evaluation. It does not contain information regarding traditional training/test/validation dataset splits, as the data is generated through interaction with the simulated environments.
Hardware Specification Yes Table 5: JAX implementation benchmark on G1-Standup, evaluated with 5 seeds on an Nvidia RTX 3090. ... Table 6: Planning performance of MTP-Akima. Averaged over 5 seeds on an Nvidia RTX 4090.
Software Dependencies No The paper mentions several software tools and frameworks: JAX (Bradbury et al., 2018), Mu Jo Co XLA (Todorov et al., 2012; Kurtz, 2024), evosax (Lange, 2023), and Gymnasium (Towers et al., 2024). However, it does not provide specific version numbers for these components, which is required for a reproducible description of ancillary software.
Experiment Setup Yes All algorithms and environments are implemented in Mu Jo Co XLA... All experiment runs are sim-to-sim evaluated... For all baselines, we fix the same number of rollouts B = 16 on Crane, and B = 128 for all other tasks. All tasks are implemented in hydrax (Kurtz, 2024). Further experiment details are in Appendix A.4. Table 2: Simulation Settings for Experiments. Table 3: MTP Hyperparameters. Table 4: PS/MPPI Hyperparameters.