Flex'N Sense: two arm policies trained in simulation (ElRobot, Isaac Lab)

Made for a hackathon only. These policies were trained by team Flex'N Sense during the Leukerbad Robotics Hackathon (November 2026) as a demonstration. They were trained and measured in simulation only. They have not been run on a real robot, they are not a medical device, and they must not be used to control a robot near a person.

The two policies

Folder Task What goes in What comes out
policies/pull_v1 Hold the top of a spring scale; give way when the finger pulls harder than 1.96 N (200 gf); come back when the pull stops 29 numbers: arm joints (angle, speed), distance to targets, the force now and one step ago, the gripper's place relative to its start, the last action 5 numbers: changes to the targets of motors 1, 2, 3, 4, 6
policies/move_v1 Take the scale near its loop, lift it over a finger, set the loop on the next finger, let go 44 numbers: arm joints, gripper command, where the gripper is and points, the grasp point, the loop and the target finger, held / placed flags, the last action 6 numbers: the same five motors, plus open / close

Each folder has policy.onnx (with policy.onnx.data), policy.pt (TorchScript), checkpoint.pt (to continue training), params/ (the exact settings) and summary.json (the measurement).

How they were trained

  • Arm: NormaCore's ElRobot, simulated from their own URDF model (commit 6b494f7, Apache-2.0).
  • Simulator: NVIDIA Isaac Sim 6.2 with Isaac Lab main@727e4e43a, 120 physics steps a second, 60 decisions a second.
  • Method: reinforcement learning, PPO from rsl_rl 5.5.1, 4,096 simulated arms in parallel on one NVIDIA RTX PRO
    1. No recording from a real arm was used.
  • Pull: 600 iterations, 59 million steps, 7.7 minutes. The scale is a spring whose stiffness, and the finger's pull (distance, speed, hold time, direction), vary per episode.
  • Move: six runs, about 470 million steps, about 80 minutes. The first five runs each exposed a flaw in the reward or in the task's rules; they were fixed before the final run. The grip in this task follows rules, not contact physics.

Measured in simulation (1,024 arms each)

Pull arm Trained policy Arm held still
Peak force, average 1.60 N 3.61 N
Peak force, worst pull 2.52 N 7.32 N
Pulls more than 10 % over the 1.96 N limit 1.1 % 85 %
Move arm
Loop placed within 8 mm of the target finger 100 %
Dropped / hit a finger 0 % / 0 %
Time to place, average 0.40 s (full motor speed: too fast for use next to a hand)

What is assumed

Servo stiffness and damping (40 N路m/rad, 2 N路m路s/rad), the spring (5.886 N over 0.10 m) and the gearbox's apparent mass on the jaws are assumptions, not measurements. The force the pull policy reads is the true simulated force plus noise; on a robot it would have to come from a camera reading of the scale. The move policy is told where the scale and the fingers are.

Files

policies/ the two policies. media/ videos of the training, including a film that goes from 1 arm to 1,024. code/ the Isaac Lab tasks and scripts that produced them. Recorded episodes of both policies are in the companion dataset repo.

Credits

The arm model is NormaCore's (Apache-2.0). The simulation uses NVIDIA Isaac Sim and Isaac Lab, and rsl_rl from the Robotic Systems Lab at ETH Zurich.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading