Camera20 Stage 2 XLeRobot EEF
a model for bimanual robot control
3 min read
A collection of fine-tuned checkpoints for the physical XLeRobot bimanual robot, designed for object manipulation tasks: sorting red and green blocks, handling fruit, and placing items into a drawer. The model takes the current state of the arms alongside frames from three RGB cameras (one head-mounted and two on the wrists) to predict a chunk of 16 future absolute target EEF (End-Effector — the 3D position and orientation of the robot's tool) poses and continuous gripper closure values. It is built on the Qwen3.5-4B architecture with 32 MoT (Mixture-of-Transformers — a multimodal architecture that separates text/vision processing from action generation) layers, fine-tuned by the ssrgroup-xlerobot team from the pretrained Camera20 Stage 2 checkpoint. Unlike a later binary-gripper revision, this release preserves smooth continuous gripper control on a 0–100 scale and uses per-task normalization. This release is aimed at researchers and engineers working on VLA (Vision-Language-Action — models connecting vision, language, and robotic control) systems.
Key features
- Three-camera vision: processes synchronized 224×224 frames from the head camera and left/right wrist cameras.
- Absolute EEF coordinates: outputs an array of shape (16, 16) looking roughly 0.53 seconds ahead (at 30 Hz) in each arm's local
base_linkframe, avoiding error accumulation from incremental steps. - Task-specific checkpoints: includes standalone weights for blocks (
green_red_blocks), fruit (15k and 30k steps), and drawers (15k steps), with no separate base Qwen download required. - EMA stabilization: all published checkpoints are saved using an exponential moving average (EMA 0.999) and 4-step flow integration.
What you can do
- Run manipulation tasks on XLeRobot: connect the policy to a bimanual robot controller to sort blocks, grasp fruit, or close drawers.
- Compare training stages: evaluate behavioral differences between intermediate 15,000-step checkpoints and final 30,000-step checkpoints.
- Use as a baseline for custom tasks: adapt the
Camera20Policyinference code and Common23 action space format to your own teleoperation setups and datasets.
Example
from camera20_inference import Camera20Policy
policy = Camera20Policy(".", "green_red_blocks")
result = policy.infer({
"state": state_eef,
"images": {"head": head_rgb, "left_wrist": left_wrist_rgb, "right_wrist": right_wrist_rgb},
})
actions = result["actions"] # (16, 16), абсолютные координаты EEFRead more
DetailsMetrics (hold-out flow loss): fruit 15k — 0.047537, fruit 30k — 0.072744, drawer 15k — 0.055822 (does not directly reflect real-world robot success rates).
Training: 30,000 steps on 36 trajectories per task, global batch size 128, q01/q99 quantile normalization.
Environment and hardware: tested on NVIDIA H100 (Python 3.10, PyTorch 2.8.0+cu128, Transformers 5.3.0, Flash Linear Attention 0.3.2, Triton 3.4.0).
Limitations: outputs are unbounded in amplitude and require an external IK/Cartesian controller; real-world rollout tests were not performed for this revision.










Comments