In an article published in the arXiv preprint server*, researchers from Switzerland developed an innovative controller called Pedipulate to enable a quadruped robot to use its legs for manipulation tasks. Their tool, trained with deep reinforcement learning, can track foot target points in a large workspace, adapt the robot’s stance and move towards far-away targets, and robustly handle external disturbances.
Figure 1:Our foot target tracking controller enables a variety of real-world manipulation tasks such as opening doors (A) and fridges (B), object transport (C), pressing a button (D), pushing obstacles out of the way (E), and collecting rock samples (F).
*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.
Background
Legged robots have potential importance in maintenance, home support, and exploration scenarios. However, most legged robots are limited to inspection work, requiring less interaction with the surroundings. To enable more complex tasks, some legged robots are equipped with a robotic arm, which increases the power consumption and mechanical complexity of the robot.
By referencing quadrupedal or four-legged animals, the researchers hypothesize that some manipulation tasks do not need the extra dexterity and complexity of robotic hands and arms; instead, they could be solved by utilizing the legs of a four-legged robot. Employing identical limbs for both locomotion and manipulation can reduce the mechanical complexity and cost of a robotic system. This advantage is especially significant in scale-limited applications like space exploration.
About the Research
In the present paper, the authors designed a reinforcement learning-based pedipulation controller by training a neural network policy that tracks foot position targets. The policy receives proprioceptive information about the robot’s state and foot position commands as input observations and outputs desired deviations from the current joint positions as actions. It was trained to minimize the distance between the foot and the target point, so designed for being close to the target point and penalized for jerky motions, joint torques, velocities, and collisions.
The study utilized Isaac Gym as a simulation environment and employed proximal policy optimization (PPO) as a deep reinforcement learning algorithm. It introduced randomization of simulation friction parameters, added random pushes to the robot’s base, and conducted training on irregular terrain to facilitate simulation-to-real transfer. Additionally, the controller was deployed on the quadrupedal robot ANYmal D, which features 12 torque-controlled joints and four feet equipped with force-torque sensors and was developed by the ANYbotics robot manufacturer.
The researchers incorporated a curriculum for adaptive command sampling, enabling the controller to approach distant targets with a raised foot using a tripod gait. Additionally, they defined the command in a fixed local control frame, allowing the user to intuitively control the foot in a fixed frame rather than having the behavior influenced by the current base pose.
Research Findings
The paper evaluated the reachable workspace and benchmarked the tracking performance of the presented pedipulation controller in simulation and on the real robot. It demonstrated that the controller could reach a large workspace from inherent whole-body movements and locomote toward faraway targets through curriculum-based command selection. Additionally, it achieved an average tracking error of 0.037 m in simulation and 0.057 m on the real robot for close-range targets and 0.043 m for far-range targets.
Furthermore, the authors assessed the performance and robustness of the proposed device in hardware experiments, showcasing various real-world manipulation skills using teleoperation. The main findings included:
- The controller could track foot target points in a large workspace, reaching a local workspace of 2.2 m x 2.7 m x 1.3 m without altering the stance configuration. It could adapt the stance or walk on three legs to reach further targets.
- Combining locomotion and manipulation (loco-pedipulation) using curriculum-based command sampling, the controller automatically switched between standing and walking modes based on the target distance. It could use a tripod gait to approach distant targets with the foot in the air for object transport.
- Remaining robust against external disturbances, such as interaction forces on the foot and base and slippery surfaces like whiteboards, the controller compensated for base motion and stance slip by adjusting joint positions and torques, adapting stance and step to maintain stability when the base was further disturbed.
- The controller achieved solely by tracking foot target points without task-specific adaptations by enabling various real-world tasks like door opening, fridge opening, backpack transport, button pressing, obstacle pushing, and rock sample collection.
Applications
The presented tool enables a broad range of manipulation tasks utilizing the legs of quadrupedal robots, which can be employed for many applications that require mobility and interaction. For example, the controller can be used for
- Maintenance tasks include inspecting and operating valves, switches, and doors in industrial sites.
- Home support tasks include fetching and carrying objects, opening and closing appliances, and pushing furniture.
- Exploration tasks like collecting and transporting samples, probing the environment, and removing obstacles in natural or planetary environments.
Conclusion
In summary, the novel tool was a versatile low-level controller that effectively tracked foot target points and was robust to disturbances. The authors highlighted that the new device could enable solving various real-world tasks via pedipulation and could be utilized for both prehensile and non-prehensile pedipulation scenarios.
The researchers acknowledged limitations and challenges and suggested that future work should move towards autonomous pedipulation using a hierarchical approach with a task-specific high-level planner and the pedipulation controller as a low-level controller. They also recommended that tracking interaction forces could extend the range of admissible manipulation tasks.
*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.
Article Revisions
- Feb 27 2024 - Main article image changed. Text change "arXiv* server" to "arXiv preprint* server"
Scaling Large Language Models Makes Them Less Reliable, Producing Confident but Incorrect Answers