Experimental · Not accepted · Separate from the default controller

SO-101 residual policy lab

step 10000 is the latest checkpoint in this training batch, not the best-performing one. All 11 checkpoints failed the original eligibility checks. This separate browser experiment does not change the current default automatic controller; the legacy ACT weights are retained.

Waiting for the simulator…

Waiting

Technical diagnostics (original JSON keys)
How it works, evidence and limits

The dedicated experimental contract uses state[1,12]: six measured joint values plus six previously applied targets, with env[1,12]. It returns absolute_target[1,6]. Normalization and residual decoding are inside the ONNX graph; the browser does not add the joint state again. The first five joint and target values are degrees; the sixth is gripper opening in percent.

A full run is 600 control steps at 30 Hz: 20 seconds of simulation time. Each control uses seven physics steps of 1/210 second. Slower devices may take longer in real time. First-frame inference does not move the simulation.

Inference, simulation and record generation run in your browser. This experiment does not upload your files or records and does not connect to a physical robot.

A strict placement result describes this one simulated run. It is not quality acceptance or physical robot validation. Missing geometry or contact evidence produces a review-needed result. Incomplete runs cannot be marked successful. Identical results across engines or devices are not guaranteed.

The original delivery contained experimental weights and failed evaluation evidence, with no ONNX model. This browser model was exported separately from the checked step 10000 weights. All 11 original checkpoints remain eligible=false and selected_step=null.

The original step 10000 development results were 18/32 with MuJoCo 3.3.7 and 15/32 with 4.0.0; the baselines were 32/32 and 28/32. The two fixed archived scenes here do not replace the development, difficult-case or independent final evaluation sets. The original feature source was not fully recovered.

Only archived seeds 35000 and 35001 are available. Preparation locks the initial state and seed. Use this panel to stop or prepare again. Changing the seed invalidates preparation; an existing record keeps its captured seed, including in its download filename.

Not verified here: a new trustworthy PyTorch rerun, the complete difficult-case and final sets, bitwise agreement across environments, physical phones or a real robot. Numerical checks do not establish grasp-and-place quality.

Loading the physics engine…

Drag to turn the view, scroll or pinch to zoom.

Open your own files

Open a folder with an MJCF model (.xml) and its meshes to load a robot of your own, STL files with the same names to swap parts of this one, or an ONNX policy to let it drive the servos. You can also drop files onto the 3D view.

Files stay in this browser. The dry run checks tensor compatibility and finite values; physical behavior remains unverified. Joint-vector spec example · SO-101 state-ACT spec example

Experiments and pose book

Record up to 60 simulation seconds or 12,000 samples. Downloads contain initial state, controls, trajectory, versions and file hashes; local model files stay on this device.

SO-101 tasks record automatically. Starting another task replaces the previous episode; download it first to keep it.

Playback sets the stored poses. Rerun requires matching engine/model/policy hashes and recomputes physics from recorded controls, without policy inference. Results need not match bit for bit across environments. The pose book lasts until this page closes.

Record an experiment or save the current pose.

Servo target angles (degrees)