
Robot Arms
SO-101
Modalix DevKit, PoLiMa with SDK 3.0 support, and an SO-101 follower with overhead and wrist USB cameras aligned to the training views.
SiMa.ai compiled this task-specific ACT policy for the Modalix DevKit. An SO-101 follower uses two USB camera views to pick up a gray eraser and place it in a white basket. The public BF16 bundle comes from step 100,000, contains six compiled graphs and predicts 100 six-dimensional actions per chunk.
View full model descriptionModalix DevKit, PoLiMa with SDK 3.0 support, and an SO-101 follower with overhead and wrist USB cameras aligned to the training views.
Paper / report unconfirmed
Official introductory materials are recorded; no version-specific paper or technical report has been verified.
Last checked: 2026-11-19
Training code for this version is unconfirmed. Public code resources do not establish access to the exact training implementation. View verification notes
SourceVerified on · 2026-11-19
The release contains compiled ELFs, reference fixtures and runtime assets. The training checkpoint and dataset are outside this distribution. Changes in geometry, lighting or calibration can affect behavior. Physical task success rate is unreported. Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.
joint_position · continuous
6 DoF · horizon 100
Single embodiment
Action control
Community implementation
100 × 6
Single task
769 demonstrations
Palette SDK 3.0.0 · Model Compiler 3.0.0 · LLiMa 0.6.0
act-so101-grey-eraser-pickplace-100k-bf16-219727d9
Modalix DevKit, PoLiMa with SDK 3.0 support, and an SO-101 follower with overhead and wrist USB cameras aligned to the training views.
SiMa.ai reports a median model latency of 13.5 ms, cosine similarity of 0.999992 and maximum absolute error of 0.0147 over 20 fixture plans after three warm-ups. The reference inputs were replayed without moving the robot.
The release contains compiled ELFs, reference fixtures and runtime assets. The training checkpoint and dataset are outside this distribution. Changes in geometry, lighting or calibration can affect behavior. Physical task success rate is unreported.