tennis_gs_env — stationary step 29

200 tennis-catching sequences in the original IsaacSim 3DGS scene and original ring appearance. All sequences have 51 frames at approximately 30 Hz and three 480×360 RGB cameras. The query is frame 29, with observation frames 23,26,29. Commands and measured robot states are verified unchanged throughout frames 0–29.

The original collection IK controller generates the demonstrations: movement starts at frame 30 and uses 15-frame smoothstep interpolation. VLA inference is run offline on every saved query; its predictions do not control these trajectories. Collection success statistics are therefore not VLA closed-loop success rates. All three cameras are used without upper-camera masking.

Per-sequence files

Each episodes/episode_XXXX/ directory contains:

• rgb_left.npz, rgb_right.npz, rgb_upper.npz: all 51 frames as lossless uint8 rgb arrays [51,360,480,3], plus frame_index.
• model_observation_step29.npz: exact raw adapter inputs; three observation.images.* arrays [3,360,480,3], observation.state [3,9], and obs_frames=[23,26,29].
• inference_request.json: task, frame indices, timestep, episode/request identifiers and policy seed.
• vla_core_inputs_step29.npz: actual processed image tensors, image masks, language token IDs/masks, normalized/padded state and initial noise passed to the action sampler. Tensor lists use keys such as images.0.
• vla_tensor_dtypes.json: original PyTorch tensor dtypes. Bfloat16 values, if present, are stored losslessly as float32 in NPZ and reconstructed using this metadata.
• vla_prediction_step29.npz: normalized and unnormalized 16-action chunks, restored terminal action, current command, and hypothetical current-command-plus-terminal-delta target. This target is not executed.
• trajectory.npz: all measured states, commanded actions, ball positions/velocities, ring centers/normals, base/ring poses and six joint velocities.
• demonstration_actions.npz: original collection relabeling action[t]=state[min(t+1,50)], plus query29 absolute and delta chunks corresponding to measured states 30–45. Angle deltas at indices 2–8 are wrapped to [-pi,pi]. Demonstration targets and model predictions are separate.
• camera_extrinsics_step29.npz: each camera's transforms at frames 23,26,29.
• labels.json: planned catch point, actual ball/ring positions at planned arrival, corresponding and final arm states, final command, IK solution, joint names, closest-approach positions, interpolated ring-plane crossing positions, stationarity checks and geometric scores.
• cameras.mp4: 51-frame left/right/upper triptych for visualization. Exact model input must use the lossless NPZ images.

Coordinate and timing conventions

Frame indices are zero-based and captured after each physics step/render. A flight of N physics steps reaches its target at stored frame N−1. The query is strictly before the first commanded movement at frame30.

Robot state order is world base X, world base Y, base yaw, then six joint angles. Positions are metres, angles radians, velocities metres/s or radians/s. Pose quaternions are XYZW. Camera matrices act on column vectors and map USD camera coordinates (+X right,+Y up,−Z forward) to Isaac world. OpenCV optical-to-USD conversion is diag(1,-1,-1,1). Extrinsics preserve the collection convention of fixed camera orientation plus measured chassis translation.

planned_catch_target_world_m is an airborne target at Z=0.9 m, not ground impact. ball_at_target_time_world_m and ring_at_target_time_world_m are measured at the planned arrival frame. Closest-approach coordinates are sampled measurements; plane_crossing_positions explicitly labels interpolated crossing coordinates. Full trajectories permit alternative event definitions.

Ring inner radius is approximately 0.086 m; ball radius 0.0325 m. Legacy success means minimum sampled center distance <=0.13 m. Whole-ball clearance means an interpolated ring-plane crossing with radial error <=0.0535 m. Geometric crossing does not demonstrate retained possession.

Model inputs and reproduction

Current checkpoint: equal_h15_3cam, SHA-256 5159343ad41b8198eba5433444ecae6bb490756fc9d7ea1ddd9059f8389b7c77. Three temporal observations use offsets [-6,-3,0]. Raw RGB becomes [1,T,C,H,W] float32 divided by255; state becomes [1,T,9]; task is a one-element list; index is [29]; dt_scale is1. The checkpoint applies its exact image, state and language preprocessing, whose resulting tensors are also saved. Inference uses ten denoising steps, noise scale1 and policy_seed=episode index. Explicit initial noise is retained.

model_metadata/ contains checkpoint configuration and hash, normalization tensors, tokenizer, camera order, runtime versions and matching inference source. Model weights are a separate prerequisite identified by the hash. load_gs_episode.py loads the arrays without IsaacSim; restore_core_arguments reconstructs tensor dtypes and sampler arguments, including noise.

from load_gs_episode import load_episode, restore_core_arguments
episode = load_episode('/path/to/downloaded/dataset', 0)
raw = episode['observation']
sampler_args = restore_core_arguments(episode, device='cuda:0')
# With the matching loaded DynamicVLA core:
# normalized_actions = core.sample_actions(**sampler_args)

First and last queries are additionally replayed from the saved core tensors/noise and compared against original sampler outputs.

Collection and index

Collection uses seed0 as one continuous RNG stream with unreachable-target rejection sampling; episode indices are not independent throw seeds. Target XY uses the original 1.4 m radius and forward-minimum0.3 m settings, target Z0.9 m, flight duration1.5–1.6 s quantized to physics_dt0.0333333, throw period1.7 s, and base landing share0.7. Full parameters, geometry and camera intrinsics are in root JSON files. The collector's original CLI arguments are preserved as provenance; this export stores the current RGB/state VLA modalities and is a custom NPZ/JSON dataset, not a drop-in LeRobot directory.

data/episodes.parquet is the episode index for Hugging Face Datasets. validation.json reports every-sequence checks and SHA256SUMS records file integrity. collection_code/ records the capture/export/inference/validation extensions; the original simulator installation and scene assets remain prerequisites for recollection.

kinematics/ contains a NumPy-only action/state-to-ring-center forward-kinematics implementation, exact USD joint-frame parameters, Chinese calculation instructions and validation against all 10,200 measured frames. It distinguishes absolute action targets, VLA delta outputs, the ring body origin and the ring geometric center.