Teleoperations

Collecting human demonstration data with the robot embodiment "in-the-loop" continues to be a core part of training Physical AI models. Anvil addresses this need through both leader-follower "ALOHA" style robot control, as well as through handheld motion tracking controllers.

VR Controller "Quest" Teleoperation

Anvil has integrated Quest Pro Controllers into its teleoperation framework, creating a robot control experience that feels natural, precise, and responsive. Anvil has gone much further than other VR controller integrations:

• Motion That Feels Natural and Stays Stable Anvil’s advanced inverse kinematics engine continuously selects arm configurations that feel natural and remain stable, even in difficult or constrained positions. It smoothly handles out-of-reach positions, prevents oscillations and sudden jumps, and maintains natural human-like arm posture for intuitive control.

• Ultra-Low Latency Control Typical VR teleoperation solutions suffer from delay, jitter, and unpredictable performance. Anvil’s Quest system uses a direct wired connection to the control computer, eliminating those sources of latency entirely.

• Teleoperation without wearing a headset Anvil’s system uses Quest Pro controllers paired with a Quest headset that can remain unworn and off to the side. Because the controllers have their own onboard cameras and a dedicated wireless link to the headset, they maintain precise tracking without needing the headset to be worn.

Leader-Follower Teleop

Anvil supports "ALOHA" style leader-follower joint-space teleoperation, and has shipped hundreds of these OpenARM systems to customers. Anvil's implementation includes gravity compensation on both the leader and follower arms, enabling easier human input on the leader, and more compliant control on the follower. All of this is executed within a single-threaded, 500 Hz control loop built in C++ on top of the ros2_control framework, resulting in ultra-low-latency control.

Many teams are very focused on leader-follower joint-space teleoperation, and as such, we do sell leader-follower systems. However, through our own testing, we've experienced leader-follower teleop to be less effective than the VR controller Quest teleop, for a couple different reasons:

• Occlusion and Visibility Challenges Having the leader and follower OpenARMs side-by-side creates an occlusion issue, where the operator has to look through the leader robot, and still can't quite see the far side of the follower robot. This creates guesswork and mental fatigue for the operator, while also resulting in slower commanded robot movements.

• High Robot Inertia Due to motor gearing and reflected inertia, the leader robot can feel much heavier than expected to an operator. Even with near perfect gravity compensation, this high inertia can be physically fatiguing to an operator, and make it harder to perform fine movements.

• Managing Additional Degrees of Freedom (e.g. Over-Articulation) The OpenARM has 7 degrees of freedom in joint space (7 joint motors) to control the 6 degree of freedom cartesian-space gripper (x, y, z, roll, pitch, yaw). Some users have experimented with constraining the 7th degree of freedom by strapping the robot elbow to the operator's elbow, but this makes it cumbersome to begin and end a data collection session, and the robot size and elbow location doesn't always match the operator's size, making the elbow-to-elbow mapping ergonomically uncomfortable for many operators as well.

Leader-Follower "GELLO" Teleop (Coming Soon)

To overcome many of the issues mentioned above with traditional leader-follower teleop, Anvil is developing a GELLO style leader-follower system for the OpenARM. Coming Soon!