From recorded state to shared-world coverage.

Coordinate frame

Positions use metres in a right-handed canonical shared-world frame. Each tool position is computed in its local follower base_link frame, then transformed as p_world = R · p_local + translation. Rotation uses Rz(yaw) · Ry(pitch) · Rx(roll).

The demo places the left base at world Y +0.4 m and the right base at world Y −0.4 m. The resulting 0.8 m separation is provisional, configurable, and not a claim about calibrated robot geometry.

Counting without ambiguity

Dataset frames

Recorded timesteps selected from the source dataset.

Tool-point visits

Left and right visits summed. A single dataset frame can contribute two tool points.

Arm-specific voxel entries

Occupied voxels retained separately per arm, even when both arms share one integer grid cell.

Unique shared grid cells

Distinct integer voxel indices after both arms enter the canonical world frame.

Exact episode identity

Each occupied arm voxel stores sorted episode identities in a compact CSR representation. Radius queries and uncommon-space scoring can therefore use exact source episode IDs rather than approximate them from aggregate counts.

Uncommon-space episode scoring

Uncommon-space exploration ranks recorded episodes by how uncommon their reached workspace entries are within this exported coverage set. It is an exploration aid, not an anomaly detector. It does not measure task quality, success, usefulness, safety, or physical novelty.

Scoring uses distinct-episode incidence from the exact CSR episode identities. Left and right arm voxel entries remain separate analytical entries, even when their integer voxel coordinates match. Raw tool-point visit counts are a separate metric and do not affect uncommonness scores.

Formula in words and notation

E

The number of episodes in the exported coverage set.

c_v

The number of distinct exported coverage episodes represented in arm-specific voxel entry v.

V_i,S

The unique arm-specific entries in scope S whose exact CSR identities contain episode i.

When E is greater than one, entry rarity is r(v) = ln(E / c_v) / ln(E). In words, this is the natural logarithm of total coverage episodes divided by the entry's distinct episode count, normalized by the natural logarithm of the total. When E is one or less, r(v) = 0.

Episode uncommonness is U(i,S) = mean of r(v) over v in V_i,S. The displayed score is 100 × U(i,S), normalized to 0–100. It is not a probability or percentile. The evidence count is displayed separately, and averaging prevents breadth or episode length alone from dominating the score.

Scope and deterministic ordering

Entire coverage is the default scope. Selected radius uses the exact arm-specific entry identities returned by the shared-world radius query. Runtime arm spacing never changes an entry's rarity, but it may change which entries geometrically fall inside the radius. Arm visibility does not change the analytical scope.

Results are ordered first by score descending, then by arm-specific entries touched descending, and finally by episode ID ascending.

Playback and limitations

Coverage scoring requires no trajectory load. Playback availability is checked lazily from the existing trajectory payload, without an additional initial request. An episode absent from that payload remains valid coverage evidence and is labelled coverage-only. The viewer does not infer a synthetic trajectory, optional state, or video. Orientation and raw gripper data match only the selected exported trajectory episode.

Scores describe only the episodes and voxelization in this exported bundle. They do not establish rarity in the full source dataset, future recordings, production behavior, or the physical workspace generally. Results depend on voxel size, selected radius, episode selection, robot model, forward kinematics, and shared-world transforms. A high score supported by few entries should be read together with its displayed touched-entry count. With one exported coverage episode, relative uncommonness is unavailable and scores are defined as zero.

Scores are derived client-side from existing coverage CSR data. No schema revision or additional initial network request was required, and work remains linear in relevant CSR membership for the current bundle. This does not claim full-dataset scale support; that remains a later roadmap phase.