Two recorded demonstrations

Two complementary studies: rigid configuration-space planning and flexible-object manipulation. Both use the original complete dual ALOHA robot geometry.

Unlocking the Claw

The puzzle geometry and object-space reference path come from the supplementary material for C-Space Tunnel Discovery for Puzzle Path Planning. This project did not discover that reference solution from scratch.

The implemented work includes local path corrections, grasp selection, full robot inverse kinematics and motion planning, placement fixtures, release and withdrawal, and geometry verification. All 3,083 motion intervals have continuous geometric checking in the recorded complete-run evidence. Support is conditional static feasibility under the stated mass and friction assumptions; it is not a dynamic settling experiment.

The parts start pregrasped. Joint timing describes the kinematic path, not a validated hardware controller. The optional panel shows saved joint positions and correctly leaves native force and control channels unavailable.

Claw playback data

Three-ring threading

The 1.1 m rope has 80 connected capsule segments and 81 nodes. The original robot, joint and actuator limits, ring geometry, supports, rope contacts, friction, and gravity remain active. Recorded acquisitions use opposing original finger-pad contact before activating the disclosed ideal point grasp. The native timestep is 0.5 ms.

The sequence includes pickup, alignment, ring insertion, changes of grasp, a real free-end retreat at the third ring, and recovery. The final raw material order is R1−, R2+, R3+. Intermediate crossing counts remain visible in the replay, including the temporary extra crossing during recovery.

The last stage has eight separate checks: continuation state and control, saved full integration states, grasp geometry and load observation, ring–rope contacts, robot contacts, material route, final release and settling, and complete arm separation from the ring apertures. These are separate verification programs with shared model dependencies, not an external laboratory or an independent second dynamics solver.

Final support uses the declared v2 protocol: every observation retains the finite-table endpoint test, a 1 mm surface-height tolerance, and all 81 node speeds below 5 mm/s. Each end region must have more than 1 mN mean signed upward table force over the preceding 20 ms. The final stable window lasts at least one second after withdrawal. Original native penetration remains below 1 mm. The first final attempt failed an instantaneous-force rule; its report is retained. A separately executed v2 run produced bitwise-identical saved physical states, with the revised observation rule explicitly declared before execution.

Threading playback data

How the viewer works

One timeline covers each full demonstration. Playback speed changes viewing time only. Claw defaults to 4× (approximately 40 seconds), and threading to 8× (approximately 47 seconds). Chapter markers are navigation aids, not success labels. The phase label and ring counts report the saved data at the current point.

Threading loads phase data as needed and retains at most three phase buffers. Turning off smooth motion shows saved poses directly. Smoothing interpolates body poses for threading and joint positions for Claw; these visual frames are never treated as physical observations. The action panel always reports the preceding saved sample, so it can differ slightly from a smoothed image.

The threading control trace has 14 original position-actuator channels: six arm channels and one gripper driver on each side. Hinge inputs are radians and gripper inputs are metres. They are not measured torque. The coupled second finger has a joint state but no separate control channel.

Rendering is capped at 30 updates per second while playing, pauses when the page is hidden, and stops when idle. The scene and controls work without a CDN, analytics, account, or live simulation backend.

Reproduction and limits

This site contains the recorded motion, geometry and actuator inputs needed for interactive viewing. Simulation source code and detailed verification reports are kept in the separate research archive. A saved-state replay is distinct from executing a planner or integrating the controller again.

Claw has recorded regeneration checks. For threading, an 8 s frozen replay was reproduced exactly; a fresh end-to-end rerun of all 51 historical phases has not been established. Early recordings lack some per-frame full integration states: 5,038 frames in 13 phases, with incomplete full-state evidence at 12 early boundaries. Later complete states cannot fill those historical gaps. One explicitly disclosed phase adds a terminal geometric observation from the original saved terminal state without changing its physics.

The work demonstrates specific planned motions. It does not establish an online policy that solves arbitrary human-adjusted initial states, frictional force closure, or operation on physical hardware.

Sources and credits

• ALOHA: MuJoCo Menagerie, pinned robot assets; copyright Trossen Robotics, BSD-3-Clause notice.
• Claw geometry and reference path: official C-Space Tunnel Discovery supplementary material. This viewer uses the original supplementary geometry. A separate geometry redistribution license has not been established. The main research repository's GPL license is not assigned to the puzzle geometry.
• Rendering: Three.js 0.180.0, MIT; original license included with the vendored files.
• Implementation and verification: the primary coding agent with auxiliary agents for planning, source review, geometry checks, and release preparation. Human input specified the task priorities, physical acceptance requirements, and presentation direction.

Project-authored viewer and implementation code use the included MIT license. Third-party files retain their own notices and terms. This page hosts the interactive viewers and their recorded data; simulation source code is not included on this site.

Return to the demonstrations ↑