Translating Robot Skills: Learning Unsupervised Skill Correspondences Across Robots
Authors: Tanmay Shankar, Yixin Lin, Aravind Rajeswaran, Vikash Kumar, Stuart Anderson, Jean Oh
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our approach s ability to learn skills correspondences across the following robots the Sawyer robot (Sawyer), the Franka Panda robot (Franka), the Baxter left hand (Bax-L), and the Baxter right hand (Bax-R). We consider domain pairs between each of these robots (Franka to Sawyer, Bax-R to Franka, Bax-L to Sawyer, Bax-R to Sawyer). In addition, we also consider translating from human demonstrators to each of the above robots. This results in 4 additional domain pairs. We compare our approach against the following alignment baselines in Table 1. |
| Researcher Affiliation | Collaboration | 1Robotics Institute, Carnegie Mellon University, Pittsburgh, PA, USA 2Meta AI Research, Pittsburgh, PA, USA. |
| Pseudocode | Yes | We present the full algorithm for learning unsupervised skill correspondences below: Algorithm 1 Translating Robot Skills |
| Open Source Code | No | The paper includes a link to a dynamic visualization of results (https://sites.google.com/view/translatingrobotskills/home) but does not provide a link or explicit statement about the release of its source code for the methodology described. |
| Open Datasets | Yes | For the Sawyer robot, we use the Roboturk dataset (Mandlekar et al., 2018), which consists of roughly 2000 demonstrations across 8 different tasks. For the Franka robot, we use the Robo Mimic dataset (Mandlekar et al., 2021), which has 800 demonstrations across 4 tasks. For the Baxter robot, we utilize the MIME dataset (Sharma et al., 2018). For human demonstrations, we consider the GRAB dataset (Taheri et al., 2020), which consists of 10 different people manipulating various objects. |
| Dataset Splits | No | The paper does not explicitly provide details about training, validation, or test dataset splits using percentages, absolute counts, or references to standard splits. It mentions using datasets but not their partitioning. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. It only mentions general concepts like "robots". |
| Software Dependencies | No | The paper mentions using "Pytorch" and "Adam optimizer" but does not specify version numbers for these or other key software components. |
| Experiment Setup | Yes | While training our translation models, we simply optimize L = Lf + Lb using the Adam optimizer (Kingma & Ba, 2014), implemented in Pytorch. We use the default parameters of Adam, i.e. a learning rate of 10 4. For our training, we use a batch size of 32. We train our translation models over 8000 epochs for each domain pair. We set the random seed for our training to 0 manually. We add in epsilon noise to our training during sampling from the learnt networks. Here, we use an initial epsilon value of 0.3, and decay the epsilon value to 0.1 over 200 epochs. |