Recorded on 23 October 2026 with the wrist-RGB + camera-points student (DAgger, 3,000 iterations), the starting checkpoint of the 7-GPU PPO run (21,504 environments). Left: external MuJoCo view for viewing only. Right: the exact 320×180 wrist frame the policy receives, rendered by the new simulator (ray-traced robot and bottle composited over the 3DGS scan, with camera effects).
Updated: the first version of these videos used an 8.2 cm camera near plane that cut the gripper fingers out of the wrist view. The near plane is now 1 cm and the fingers render fully, as on the real camera. This student was trained on the clipped view, so it does worse on the corrected one: 5 of 8 recorded episodes succeeded (7 of 8 before); the PPO run was restarted with the fix. Three are shown, including a failure. These are illustrations, not a success-rate estimate.
Policy input wrist RGB + history + proprio + camera pointsActions 25 Hz joint targets · binary gripperBottle scale 1.00 · eval bank, normal starts
Upright bottle · success in 4.4 s · external view | policy wrist cameraOpen video ↗Lying bottle · success in 4.1 s · external view | policy wrist cameraOpen video ↗Lying bottle · failure (timed out at 20 s) · external view | policy wrist cameraOpen video ↗
Deterministic actions, task success rule (lift and one-second hold), registered MJLab / MuJoCo-Warp physics on an A100. No real hardware was operated. Episode metadata (all 8 episodes, checkpoint hash).
Latest downloaded baseline checkpoint: 901 training updates. The point-cloud policy picked a lying bottle from an ordinary initial state and completed the stable-lift test in 10.2 seconds, including the one-second hold. This successful example is not a success-rate estimate.
Actions 10 Hz relative EE · binary gripperTraining SGS · 32,768 environmentsPhysics overflow None recorded
External simulation view · normal-speed playbackOpen video ↗
Tool-frame translation and rotation deltas are relative to the measured EE pose, bounded to 1.6 cm and 5° per action. Targets are held for 100 ms and tracked at 50 Hz. Binary gripper; no scripted arm pause. Recorded locally with registered MuJoCo-Warp physics and CPU policy inference. No real hardware was operated. Episode metadata and exact action contract.
Recorded locally on 18 October 2026 using the validated 1,000-update point-cloud checkpoint. The policy picked a lying bottle from an ordinary initial state and completed the stable-lift test in 4.44 seconds, including a one-second hold. This selected episode is an illustration, not a success-rate estimate.
External view · recorded simulation states · normal-speed playbackOpen video ↗
Six arm actions set relative joint targets at 0.04 rad per unit; the gripper commands open or close. No scripted arm pause. The policy receives scene point clouds and robot history; this external rendering is for viewing. CPU inference and registered MuJoCo-Warp simulation on the local P620. No real hardware was operated. Episode metadata and action contract.
All 11 recordings from today, including one zero-action diagnostic, with external and wrist views.
Recordings are listed newest first and labeled using their logged stop reasons. The latest three ended with stale IK packet faults. These recordings do not establish a success rate. The camera players have separate time origins. Recording details.
CONTROLLER HOLD
RGB policy · 16:06 IST
Recorded termination: fault: Stale IK packet; holding robot.
Policy bottle-rl-v36-20Action scale 0.15Time to stop 15.7 s
Recorded locally on 13 October 2026. The policy picked an upright bottle, lifted it and completed the required hold. This is the first full local episode after a two-step setup check, from an ordinary randomized start. It is one example, not a local success-rate evaluation.
Result Valid successSimulation time 8.6 sBottle clearance 16.0 cmHold 1.05 s
Synchronized external view and recorded wrist image · 1× simulation timeOpen video directly ↗Policy and local simulation details
The V26 RGB checkpoint previously scored 201/256 (78.5%) in the A100 nominal evaluation; it has not passed the final nominal-and-every-size qualification. This video uses that preserved checkpoint, not a newly qualified policy.
The policy receives 16 wrist images and robot/command histories spanning 1.5 seconds. It predicts 16 actions at 10 Hz and executes four before replanning every 0.4 seconds. Inference ran on this PC’s CPU; the Quadro P620 ran the registered MuJoCo-Warp task and 3DGS camera. The full episode took approximately 59 seconds of wall time.
The external view is rendered from recorded simulation states with the 3DGS room background; the wrist view shows the recorded policy-camera images, including visual randomization. The recording ends at success, before reset or gripper opening. No real hardware was used.
Seed 72001 · ordinary reset 192 · 173 recorded frames at 20 Hz · Rollout metadata.
Real robot · 7 October 2026
Two successful attempts and one failed attempt, shown from the external and wrist cameras.
3 selected attempts2 camera views per attemptSimulation-only policy training
These are selected clips, not a complete evaluation. The source session ran four attempts: two automatic successes, one timeout, and one workspace-limit stop. Six further planned attempts did not run.
The policy grasped and lifted the bottle. Fresh gripper feedback and the lift threshold remained valid for the one-second hold; the loop then opened the gripper and returned Home.
The policy completed another pick after the bottle was repositioned. The automatic success check passed and the loop continued through release and Home.
The attempt reached the 50-second timeout without confirming a held bottle. At the end, measured gripper width was 28.6 mm versus a 28.84 mm command, with no grip flag. Tool motion alone was not counted as a successful pick.
Policy time 50.0 sConfirmed tool rise Not confirmed
The deployed policy uses the current wrist image and 16 state observations (current plus 15 past states) spanning 1.5 seconds. Actions are spaced at 10 Hz, with a 16-action prediction horizon and fixed replanning every 0.4 seconds.
The two success clips passed the gripper-flag check, a 10 cm tool-rise threshold, and a one-second hold. The software force/contact guard was disabled for this supervised session; robot protective stops and workspace checks remained active.
External clips include a short lead-in and the end of the attempt. Wrist clips come from the individual policy recordings. The two players are independent and do not share an exact time origin. These videos are real-robot tests, not training demonstrations.