The results provide stronger evidence that RL-only bootstrapping can create the first usable language-conditioned controller for a genuinely novel embodiment, and use reinforcement learning in simulation with dense geometric rewards computed from simulator state.
Abstract
Adapting a pretrained vision-language-action (VLA) policy to a new robot usually assumes embodiment-specific demonstrations. This assumption is especially restrictive for custom robots whose morphology differs strongly from the manipulators seen in large robot datasets. We study a harder setting: zero-demo embodiment alignment of OpenVLA-OFT on a cable-driven parallel robot (CDPR) with a simple gripper and a previously unseen control interface. Instead of supervised fine-tuning, we use reinforcement learning in simulation with dense geometric rewards computed from simulator state. The training is performed in two stages: a PPO stage for directional motion primitives, followed by GRPO continuation from the PPO checkpoint with an expanded instruction space that includes object-conditioned commands. On the four shared directional instructions, the average held-out success rate improves from 34.25\% after PPO to 53.50\% after PPO$\rightarrow$GRPO, with especially large gains on \texttt{move left} and \texttt{move backward}. In the GRPO stage we additionally introduce \texttt{move to} over eight target objects and obtain 39/400 = 9.75\% strict success, while qualitative rollouts frequently show correct target-directed approach behavior before late-stage instability. Compared with prior OpenVLA and OpenVLA-OFT results, which rely on demonstration datasets and mostly standard rigid-arm embodiments, our method uses no embodiment-specific dataset at all. The results do not yet establish robust manipulation, but they provide stronger evidence that RL-only bootstrapping can create the first usable language-conditioned controller for a genuinely novel embodiment.
Two systematic attempts to improve large pretrained models with minimal or zero modification to their weights via reinforcement learning on a frozen OpenVLA-7B using binary task-success rewards on LIBERO-Goal reveal a common ceiling.
The Universal Manipulation Interface (UMI), originally developed by the Robotics and Embodied AI Lab at Stanford University, has demonstrated remarkable effectiveness for training manipulation policies for terrestrial robotic manipulators using imitation and diffusion based learning techniques. The long term objective of our research is to extend this technology to space robotic applications. There are numerous challenges largely unexplored, including harsh environmental conditions, stringent power and computational constraints, communication latency, and very limited opportunities for data collection and validation. This paper presents the first step toward achieving that goal by designing a similar gripper and testing it in both simulation and experiment with a Franka Emika robotic arm in our lab setting. We reproduce the data collection and diffusion policy training pipeline on commodity hardware, with a ViT-B/16 Vision Transformer serving as the policy’s vision backbone, and reconstruct a Franka Emika manipulator equipped with a custom electric gripper inside NVIDIA Isaac Sim, using the Lula inverse kinematics solver to perform kinematic control from the policy generated end effector commands. We identify and formalize the coordinate and action frame transformations required to transfer a policy trained on handheld demonstrations onto a simulated embodiment, and show that the simulated controller tracks the commanded trajectories to subcentimeter accuracy. We further report pick and place rollout statistics across randomized object configurations and identify the visual domain gap between rendered and real observations as the dominant remaining barrier to transfer. These results helped us understand the UMI framework and established a solid foundation for us to move forward toward free floating microgravity manipulation and autonomous dual arm object handover.
Neal D'Andrea, Joshua Wachs, Abdou Wade et al.· National Aerospace and Elect...· 0 citations
Moving large objects, such as furniture or appliances, is a critical capability for robots operating in human environments. This task presents unique challenges, including whole-body coordination to avoid collisions and managing the underactuated dynamics of bulky, heavy objects. In this work, we present RobotMover, a complete learning-based system for large-object manipulation that leverages human–object interaction demonstrations to train robot control policies. RobotMover formulates the manipulation problem as imitation learning using a simplified spatial representation—referred to as the interaction chain—to capture essential human–object interaction dynamics in a morphology-agnostic way. We integrate this interaction chain into a reward structure and train policies in simulation using domain randomization to support zero-shot transfer to real-world hardware. The learned policies enable a spot robot to manipulate various large objects—including chairs, tables, and standing lamp. Through extensive experiments across simulation and real-world platforms, we demonstrate that RobotMover achieves strong performance in terms of capability, robustness, and controllability, outperforming both learned and teleoperation baselines. Our system further supports practical applications by combining the learned policy with simple planning modules to accomplish long-horizon object transport and rearrangement tasks in the real world.
Tian-Yu Li, Joanne Truong, Tsung-Yen Yang et al.· IEEE Transactions on robotic...· 2 citations
This short paper revisits the Push-T task in the context of emerging advances in Agentic Robotics where an LLM coding agent -- Claude Code with Fable 5 -- is prompted to create an algorithmic solution that does not require any demonstration data.
Shuang-Yu Xie, Kai-Peng Chen, Ken Goldberg· 0 citations
This work uses Sample-based Model Predictive Control entirely in simulation as an automated, rapidly tunable expert to generate massive offline datasets and validate the robustness of this sim-to-real framework by successfully deploying complex loco-manipulation skills across different morphologies.
Martin Schuck, Maks Sorokin, S. Manni et al.· 0 citations
Simulation Pre-training for Dexterity (SPD) is introduced, a pre-training framework for dexterous manipulation that uses data entirely collected in simulation and outperforms training behavior cloning policies from scratch, showing that simulation teleoperation is a viable pre-training source for real-world dexterous manipulation.
Sarthak Kamat, Adam Rashid, Satvik Sharma et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
