feat: add tfds dataset format and bridge v2 dataset

neuracore-dictionary.txt

@@ -250,8 +250,11 @@ nans
 posinf
 neginf
 rotvec
+roundoff
 dataconfig
 pbtxt
+tfrecord
+widowx
 calcsize
 EADDRINUSE
 metas
@@ -270,4 +273,4 @@ CREAT
 getpid
 WRONLY
 rels
-huggingface
+huggingface

neuracore/importer/config/bridge_v2.yaml

@@ -0,0 +1,78 @@
+input_dataset_name: bridge_v2_dataset_converted_externally_to_tfds
+
+output_dataset:
+  name: bridge_v2
+  tags: [widowx, bridge_v2, manipulation]
+  description: "BridgeData V2 is a large and diverse dataset of robotic manipulation 
+  behaviors designed to facilitate research in scalable robot learning. The dataset is 
+  compatible with open-vocabulary, multi-task learning methods conditioned on goal 
+  images or natural language instructions. Skills learned from the data generalize to 
+  novel objects and environments, as well as across institutions.
+
+. Dataset source: https://rail-berkeley.github.io/bridgedata/"
+
+robot: 
+  name: widowx
+  urdf_path: "neuracore/neuracore/importer/config/widowx.urdf"
+  override_existing: true
+
+frequency: 5.0
+
+data_import_config:
+  RGB_IMAGES:
+    source: observation
+    image_convention: CHANNELS_LAST
+    order_of_channels: RGB
+    mapping: 
+      - name: image_0
+        source_name: image_0
+      - name: image_1
+        source_name: image_1
+      - name: image_2
+        source_name: image_2
+      - name: image_3
+        source_name: image_3
+
+  JOINT_POSITIONS: 
+    source: observation.state
+    units: RADIANS
+    mapping:
+      - name: joint_1
+        index: 0
+      - name: joint_2
+        index: 1
+      - name: joint_3
+        index: 2
+      - name: joint_4
+        index: 3
+      - name: joint_5
+        index: 4
+      - name: gripper_revolute_joint
+        index: 5
+
+
+  PARALLEL_GRIPPER_OPEN_AMOUNTS:
+    source: observation.state
+    inverted: false
+    mapping:
+      - name: gripper_open_amount
+        index: 6
+
+  VISUAL_JOINT_POSITIONS:
+    source: observation.state # path to the data in the input dataset
+    format:
+      visual_joint_type: GRIPPER
+    mapping:
+      - name: gripper_prismatic_joint_1
+        index: 6
+        inverted: false
+        offset: 0.0
+      - name: gripper_prismatic_joint_2
+        index: 6
+
+  LANGUAGE:
+    source: language_instruction
+    format:
+      language_type: BYTES
+    mapping:
+      - name: instruction

neuracore/importer/core/validation.py

@@ -25,6 +25,10 @@
     if DATA_TYPE_TO_NC_DATA_CLASS[dt] == JointData
 ]
 
+# Small absolute tolerance for floating point comparisons against joint limits.
+# This avoids false-positive warnings caused by float32/float64 roundoff.
+JOINT_LIMIT_EPSILON = 1e-6
+
 
 def validate_rgb_images(data: Any, format: DataFormat) -> None:
     """Validate RGB image data.
@@ -135,17 +139,13 @@ def validate_joint_positions(
     if name not in joint_info:
         raise DataValidationError("Joint not found in robot model.")
 
-    if (
-        joint_info[name].limits.lower is not None
-        and data < joint_info[name].limits.lower
-    ):
+    lower_limit = joint_info[name].limits.lower
+    if lower_limit is not None and data < lower_limit - JOINT_LIMIT_EPSILON:
         raise DataValidationWarning(
-            f"Position {data} is below the lower limit {joint_info[name].limits.lower}."
+            f"Position {data} is below the lower limit {lower_limit}."
         )
-    if (
-        joint_info[name].limits.upper is not None
-        and data > joint_info[name].limits.upper
-    ):
+    upper_limit = joint_info[name].limits.upper
+    if upper_limit is not None and data > upper_limit + JOINT_LIMIT_EPSILON:
         raise DataValidationWarning(
             f"Position {data} is above the upper limit {joint_info[name].limits.upper}."
         )

neuracore/importer/importer.py

@@ -28,7 +28,10 @@
     validate_dataset_config_against_robot_model,
 )
 from neuracore.importer.lerobot_importer import LeRobotDatasetImporter
-from neuracore.importer.rlds_importer import RLDSDatasetImporter
+from neuracore.importer.rlds_tfds_importer import (
+    RLDSDatasetImporter,
+    TFDSDatasetImporter,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -226,10 +229,20 @@ def _run_import(
 
     logger.info("Setup complete; beginning import.")
 
-    importer: RLDSDatasetImporter | LeRobotDatasetImporter | None = None
     skip_on_error = args.skip_on_error
+    importer: TFDSDatasetImporter | RLDSDatasetImporter | LeRobotDatasetImporter
     if dataset_type == DatasetTypeConfig.TFDS:
-        raise NotImplementedError("TFDS import not yet implemented.")
+        logger.info("Starting TFDS dataset import from %s", args.dataset_dir)
+        importer = TFDSDatasetImporter(
+            input_dataset_name=dataconfig.input_dataset_name,
+            output_dataset_name=dataconfig.output_dataset.name,
+            dataset_dir=args.dataset_dir,
+            dataset_config=dataconfig,
+            joint_info=robot.joint_info,
+            dry_run=args.dry_run,
+            suppress_warnings=args.no_validation_warnings,
+        )
+        importer.import_all()
     elif dataset_type == DatasetTypeConfig.RLDS:
         logger.info("Starting RLDS dataset import from %s", args.dataset_dir)
         importer = RLDSDatasetImporter(
@@ -256,14 +269,16 @@ def _run_import(
             skip_on_error=skip_on_error,
         )
         importer.import_all()
+    else:
+        raise DatasetOperationError(f"Unsupported dataset type: {dataset_type}")
 
     logger.info("Finished importing dataset.")
 
 
 def main() -> None:
-    """Delegate to the Typer CLI app located in neuracore.importer.CLI.app."""
+    """Delegate to the Typer CLI app located in neuracore.importer.cli.app."""
     # Import locally to keep importer.py free of Typer dependency for library use
-    from neuracore.importer.CLI.app import main as cli_main
+    from neuracore.importer.cli.app import main as cli_main
 
     cli_main()