ajaysri/episodic_box_search_robot_sam3_v1

Episodic box search: robot demonstrations with SAM3 home points

45 bimanual ARX robot sessions, exported as 638 primitive episodes and 118,299 frames at 30 Hz. This is a derived LeRobot v3 dataset for a visual-goal-conditioned low-level policy. All retained sessions belong to the training split; no held-out evaluation performance is claimed.

The robot opens visually similar boxes, shows their contents to the overhead camera, returns boxes to their home positions, and empties requested objects into a tray. Sessions contain one or two successive object queries. The low-level primitives are open, retrieve, and close. The original event buttons confirmed completed actions, so each primitive labels the interval preceding its confirmation.

Model inputs

Three clean current RGB views (video_overhead_raw, video_left, video_right) and one separate goal_reference video bank, all letterboxed to 448 x 448. The goal is sampled uniformly from the current primitive's start through the current frame (inclusive), independently on every training access. Its overhead background comes from that sampled frame; the blue home-position dot with white border stays fixed throughout opening, retrieval, and closing. Sampling cannot cross into an earlier primitive or use a future frame. The tray has no pose marker. Task text specifies the primitive; hidden object locations and future observations are not model inputs.

observation.images.goal_reference stores a marked reference for every frame. The runtime sampler produces observation.images.goal for the fourth model camera; this output is generated dynamically, not stored as a parquet feature. Primitive episodes define the sampling boundary. The sampler uses NumPy default_rng(7). metadata.goal_reference_frame records where the fixed home geometry originated; the actual sampled primitive-local frame is observation.goal.anchor_frame at runtime.

This revision supersedes the original fixed-background export. The original remains available at HF commit 24ffca41ae6bf5c18a7897f6f86c9211bf810e85; any legacy fixed-goal video files retained in repository history/current storage are not referenced by this revision's feature metadata.

Actions and timing

State and action vectors have 14 absolute joint/gripper values, ordered left arm six joints plus gripper, then right arm six joints plus gripper. Source actions were measured robot positions, not commanded targets. Exported action[t] is the next recorded position, clamped at the current primitive's final sample. Future action chunks must clamp within the primitive. At 30 Hz, a 100-action chunk spans approximately 3.33 seconds. This export does not resample to 100 Hz.

Cameras were sampled together from the latest available streams; the cameras were not hardware synchronized. Independent camera timestamp origins are not evidence of physical image lag. Source frame and query indices are retained as metadata.

Geometry and selection

SAM3 detects resting boxes; logical identities are matched injectively to initial placement dots and updated causally through later resting frames. Points are latched before manipulating each selected box. Seven primitive intervals from three manipulations were omitted because their home points remained unresolved. Retained point estimates are automatic annotations, not exhaustive human-verified ground truth. The source recordings are unchanged.

annotations/primitives.json records source intervals, goal reference frames, box identities, and exclusions. contract.json records the derived revision, artifact hashes, conversion provenance, representation, and training recipe. Source IDs are annotation metadata, not policy inputs. For downstream HLP experiments, split by whole original session and retry family, not primitive or frame.

Intended training recipe

Pi0.7 full fine-tuning from its pretrained base; 16,000 optimizer steps, global batch 64, checkpoints every 2,000 steps, action horizon 100, and training-time RTC with uniformly sampled delays of 0-20 action steps. This card describes the training setup, not a completed or evaluated model.