One reel per row of the paper's three simulation tables (main result, data effect vs. trigger effect, policy classes), for both tasks. Square policies observe the wrist camera only (no proprioception, no fixed camera); MugCleanup policies observe the wrist camera plus proprioception (end-effector pose and gripper opening), identically in every arm of a task. The large third-person view is for the viewer only; the inset is the wrist image the policy sees. The green pill shows who controls the robot: the policy, or the primitive the policy has triggered (the perching waypoint on Square, the carry routine on MugCleanup). On MugCleanup the orange thumbnail is the wrist image at the moment the policy pressed the anchor, the pose the carry routine later runs relative to. Unassisted and no-channel policies have no primitive to trigger and run alone.

How to read the reels. Each reel is 12 consecutive seeded episodes (evaluation seeds 5000–5011, the same initial states in every condition of a task) from one training seed (seed 1), evaluated at the final checkpoint and played at 2× speed; no episode was selected or removed, and the squares under the picture fill in with each outcome. The success rates in the paper are pooled over 3 training seeds × 500 episodes per condition. Rollouts are not reproducible episode by episode, so the 12 outcomes of a reel do not reproduce the table; each caption gives the paper's pooled number next to what the reel shows.

Unassisted vs. TAP demonstrations, side by side (Diffusion Policy)

Square: main result and data effect vs. trigger effect (Diffusion Policy)

Rows of the main-result table and of the data-vs-trigger table. “No channel”: trained on the TAP demonstrations without the primitive channel; the policy runs alone and must reproduce the perch move from the wrist image.

Square: TAPs across policy classes

BC is a feed-forward MLP, BC-Transformer a causal Transformer, both with mean-squared-error heads; same demonstrations, split, seeds, image encoder, budget and evaluation as Diffusion Policy. The channel is decoded by the same rule.

MugCleanup: main result and data effect vs. trigger effect (Diffusion Policy)

“No channel”: trained on the TAP demonstrations without the channel; the policy must imitate the carry itself. The last reel is the gate diagnostic of the paper: the same TAP policy with the gripper gate on the carry press opened wide (0.317 against 0.323 with the gate).

MugCleanup: TAPs across policy classes

Same two baselines, same protocol. BC sits near the floor on this task in both arms (37 and 111 successes of 1500 episodes in the paper), so most of its episodes here fail.

Every evaluation trial

The complete evaluation set behind the tables, re-recorded with video: for every condition and training seed all 500 evaluation trials in order, eight at a time at 4× speed. Each tile shows its trial number and evaluation seed (trial t starts from the initial state of evaluation seed 5000 + t, the same in every condition of a task) and is stamped success or fail with the evaluator's verdict. Nothing is selected, skipped or reordered. Rollouts are not bit-reproducible (GPU convolution autotuning), so a recording's count can differ slightly from the paper's own evaluation run of the same checkpoint.

Square, Diffusion Policy: unassisted

wrist camera only. These recordings: 181 of 500 trials succeed (0.362). Paper: 0.372, pooled over 3 training seeds × 500 episodes.

Square, Diffusion Policy: TAP + channel

wrist camera only. These recordings: 267 of 500 trials succeed (0.534). Paper: 0.531, pooled over 3 training seeds × 500 episodes.

Square, Diffusion Policy: TAP data, no channel

wrist camera only. These recordings: 143 of 500 trials succeed (0.286). Paper: 0.297, pooled over 3 training seeds × 500 episodes.

Square, BC: unassisted

wrist camera only. These recordings: 44 of 500 trials succeed (0.088). Paper: 0.085, pooled over 3 training seeds × 500 episodes.

Square, BC: TAP + channel

wrist camera only. These recordings: 77 of 500 trials succeed (0.154). Paper: 0.250, pooled over 3 training seeds × 500 episodes.

Square, BC-Transformer: unassisted

wrist camera only. These recordings: 149 of 500 trials succeed (0.298). Paper: 0.263, pooled over 3 training seeds × 500 episodes.

Square, BC-Transformer: TAP + channel

wrist camera only. These recordings: 148 of 500 trials succeed (0.296). Paper: 0.329, pooled over 3 training seeds × 500 episodes.

MugCleanup, Diffusion Policy: unassisted

wrist camera + proprioception. These recordings: 124 of 500 trials succeed (0.248). Paper: 0.203, pooled over 3 training seeds × 500 episodes.

MugCleanup, Diffusion Policy: TAP + channel

wrist camera + proprioception. These recordings: 162 of 500 trials succeed (0.324). Paper: 0.323, pooled over 3 training seeds × 500 episodes.

MugCleanup, Diffusion Policy: TAP data, no channel

wrist camera + proprioception. These recordings: 116 of 500 trials succeed (0.232). Paper: 0.243, pooled over 3 training seeds × 500 episodes.

MugCleanup, Diffusion Policy: TAP + channel, gripper gate opened wide (diagnostic)

wrist camera + proprioception. These recordings: 171 of 500 trials succeed (0.342). Paper: 0.317, pooled over 3 training seeds × 500 episodes.

MugCleanup, BC: unassisted

wrist camera + proprioception. These recordings: 14 of 500 trials succeed (0.028). Paper: 0.025, pooled over 3 training seeds × 500 episodes.

MugCleanup, BC: TAP + channel

wrist camera + proprioception. These recordings: 68 of 500 trials succeed (0.136). Paper: 0.074, pooled over 3 training seeds × 500 episodes.

MugCleanup, BC-Transformer: unassisted

wrist camera + proprioception. These recordings: 37 of 500 trials succeed (0.074). Paper: 0.106, pooled over 3 training seeds × 500 episodes.

MugCleanup, BC-Transformer: TAP + channel

wrist camera + proprioception. These recordings: 68 of 500 trials succeed (0.136). Paper: 0.153, pooled over 3 training seeds × 500 episodes.