yuecideng · Oct 20, 2026 · Oct 20, 2026 · Oct 20, 2026 · Oct 20, 2026 · Oct 20, 2026
diff --git a/agent_context/MAP.yaml b/agent_context/MAP.yaml
@@ -272,7 +272,8 @@ topics:
   title: Differentiable Learning Layers
   aliases: [learning layers, differentiable IK layer, 可微 IK, 训练层, 学习模块分层]
   keywords: [DifferentiableIKLayer, KinematicsLayer, SerialChainKinematics, DifferentiableIKCfg,
-    solve_dls_ik, damped_least_squares_step, Jacobian gradient, dJ/dq, learn_grasp_yaw, grasp yaw, 抓取角度优化, 抓取姿态学习]
+    solve_dls_ik, damped_least_squares_step, Jacobian gradient, dJ/dq, learn_grasp_yaw, grasp yaw, 抓取角度优化, 抓取姿态学习,
+    ResNet10Encoder, ResNet10, visual encoder, 视觉编码器]
   paths: [topics/learning-layers/learning-layers.md]
   source_of_truth:
   - embodichain/compute/kinematics/dls.py
@@ -442,7 +443,7 @@ topics:
   status: active
 - id: data-pipeline
   title: Data Pipeline
-  aliases: [online data, online dataset, async recorder, dataset persistence, simple BC demo, ODS documentation, ODS 文档, 数据管线, 在线数据, 数据集持久化, 数据保存,
+  aliases: [online data, online dataset, async recorder, dataset persistence, simple BC demo, ODS documentation, ODS capability showcase, ODS 能力展示, ODS 文档, 数据管线, 在线数据, 数据集持久化, 数据保存,
     深度视频, 数据集校验, 数据集转换, 数据集划分]
   keywords: [OnlineDataEngine, OnlineDataset, sample_batch, sample_batch_with_context, lock_index, LeRobotRecorder, AsyncLeRobotRecorder,
     DepthSidecarManager, depth_meta.json, fragment_id, finalize, expert action schema, qpos_slice, qvel_slice,
@@ -468,6 +469,7 @@ topics:
   - docs/source/features/online_data.md
   - docs/source/features/ods/
   - docs/architecture/dataset-loading-design.md
+  - docs/architecture/ods-data-stream-showcase-design.md
   - embodichain/lab/gym/envs/managers/dataset_manager.py
   - embodichain/lab/gym/envs/managers/datasets.py
   - embodichain/lab/gym/envs/managers/async_datasets.py
@@ -479,9 +481,10 @@ topics:
   - tools/lerobot_export/
   - embodichain/lab/scripts/run_task.py
   - examples/ods/simple_bc_demo/
+  - examples/ods/data_stream_demo/
   watch_paths: [embodichain/data_pipeline/, embodichain/lab/gym/envs/managers/dataset_manager.py, embodichain/lab/gym/envs/managers/datasets.py,
     embodichain/lab/gym/envs/managers/async_datasets.py, embodichain/lab/gym/envs/embodied_env.py, embodichain/lab/scripts/run_task.py,
-    tests/data_pipeline/, embodichain/cli/dataset.py, tools/lerobot_export/, examples/ods/simple_bc_demo/,
+    tests/data_pipeline/, embodichain/cli/dataset.py, tools/lerobot_export/, examples/ods/simple_bc_demo/, examples/ods/data_stream_demo/,
     docs/source/features/online_data.md, docs/source/features/ods/, docs/architecture/dataset-loading-design.md]
   related_topics: [manager-functor, env-framework, task-programs, sensor-system]
   status: active

diff --git a/agent_context/topics/data-pipeline/data-pipeline.md b/agent_context/topics/data-pipeline/data-pipeline.md
@@ -16,6 +16,9 @@ belongs to [data-assets](../data-assets/data-assets.md).
 | Shared sequence contract and sources | `embodichain/data_pipeline/datasets/sequence.py`, `sources.py`, `_sampling.py` |
 | Bounded numeric LeRobot loading | `embodichain/data_pipeline/datasets/lerobot.py`: `LeRobotDataSource`, explicit field mapping |
 | Local Franka ChunkBC demo and offline comparisons | `examples/ods/simple_bc_demo/train.py`; local task/config/model/trainer/evaluator and evidence in the same directory |
+| Pure delivery capability and standard video decoding comparison | `examples/ods/data_stream_demo/run.py`, `suite.py`, `report.py`; no learner, real SDK/DataLoader decode and parent-owned prefetch quotas |
+| Fixed-budget visual training, diversity and total pipeline time | `examples/ods/data_stream_demo/train.py`, `training_metrics.py`, `training_report.py`; single/dual-camera BC, standard dataset recording and completed-training accounting; `training_throughput.py` owns three-stage preflight guidance |
+| Dual-camera learning and camera review | `examples/ods/data_stream_demo/converge.py`, `preview.py`; `CONVERGENCE.md` owns the candidate recipe, development selection and one held-out test |
 | Commit/discard/finalization dispatch | `embodichain/lab/gym/envs/managers/dataset_manager.py` |
 | Synchronous LeRobot persistence | `embodichain/lab/gym/envs/managers/datasets.py`: `LeRobotRecorder` |
 | Background persistence | `embodichain/lab/gym/envs/managers/async_datasets.py`: `AsyncLeRobotRecorder` |
@@ -38,6 +41,34 @@ Example-local documentation owns its task binding, CLI/artifact names,
 comparison protocol and historical evidence. Keep generic operational contracts
 in the core guides.
 
+The [ODS capability-showcase design](../../../docs/architecture/ods-data-stream-showcase-design.md)
+owns the fixed-budget single/dual-camera BC comparison and the historical
+state/single/dual delivery diagnostic. Its example-local decoded reader is
+separate from the public resident numeric reader. Follow it for offline corpus
+sizes, source-reuse protection, camera qualification and timing boundaries.
+Training comparisons match actual updates and valid supervision; data volume
+and trained-state diversity are outcomes. Record only completed optimizer
+updates as training coverage, and include recording/finalization and normal
+teardown in total pipeline time. Serial trace replay separately attributes
+generation/training overlap.
+Budget-aware prefetch and three-stage capacity guidance remain example-local;
+they do not change public engine/controller scheduling. The
+[visual demo README](../../../examples/ods/data_stream_demo/README.md)
+owns calibration and pilot commands, including isolated eager controls.
+The visual runner can ingest complete producer subbatches before generation
+completion. Its raw training/archive/serial owners share one C-episode bound;
+serial controls still archive during generation. This example-local handoff
+and nested SDK-stage profiling preserve the public sampling contract.
+All acquisition controls may overlap recording and generation within the same
+storage bounds. Worker SDK service timestamps remain separate from owner
+copy/wait spans. `training_audit.py` combines full numeric checks, source-window
+SDK reads and `video_integrity.py`'s all-frame software video count/integrity
+check; fresh post-timing receipts preserve earlier evidence.
+The [experiment evidence index](../../../examples/ods/data_stream_demo/EXPERIMENTS.md)
+links immutable historical studies and their workload-specific limitations.
+Generated run evidence is excluded from the source tree; it does not establish
+training efficiency or policy performance.
+
 `BatchSource.sample_context()` returns owned TensorDicts under a shared
 `SequenceSpecCfg`; `SequenceDataset` supplies the common iteration interface.
 The package's `schema.py` owns common semantic validation and schema fingerprints;

diff --git a/agent_context/topics/learning-layers/learning-layers.md b/agent_context/topics/learning-layers/learning-layers.md
@@ -8,8 +8,12 @@ These modules do not import learning or simulation. `SerialChainKinematics`
 returns TCP poses and geometric Jacobians from one FK pass, retaining dJ/dq.
 
 `learning/layers/` owns `KinematicsLayer` and `DifferentiableIKLayer` Torch
-module interfaces. `learning/models/` owns general MLPs, feature normalization
-and configuration-driven MLP construction. These components do not import RL,
+module interfaces. `learning/models/` owns general MLPs, feature normalization,
+configuration-driven MLP construction, and the trainable `ResNet10Encoder`.
+The encoder is imported lazily so torchvision is required only by visual clients;
+it consumes float32 NCHW RGB in `[0, 1]` and owns ImageNet normalization.
+The ODS visual demo owns the shared-view/history BC head and task losses, not
+the general encoder. These components do not import RL,
 TensorDict or simulation. RL policy interfaces, actor-critic composition and
 policy registration/builders belong to `learning/rl/policies/`; RL code consumes
 the shared networks and normalizer. See [RL learning](../rl-learning/rl-learning.md)
@@ -57,7 +61,11 @@ part of the training/evaluation entry point. The demo does not execute physical
 motion; keep task predictors out of production policy registries until a reusable
 contract exists.
 
-Tests live under `tests/compute/kinematics/` and `tests/learning/layers/`.
+Tests live under `tests/compute/kinematics/`, `tests/learning/layers/`, and
+`tests/learning/models/`. The visual encoder tests verify reference-structure
+forward/gradient equivalence, image ownership and checkpoint round trips;
+the ODS demo's `qualify_encoder.py` additionally verifies pinned RLInf classes,
+causal multiview inputs, image-only fitting, and explicitly selected CUDA.
 Validate FK, dJ/dq, full IK and recurrent-state gradients separately; include
 nonzero TCP offsets, repeated solves, checkpoint migration and import boundaries. GPU validation runs serially with `--run-gpu`.
 The unified kinematic benchmark's `differentiable` mode measures training and

diff --git a/docs/architecture/dataset-loading-design.md b/docs/architecture/dataset-loading-design.md
@@ -9,6 +9,12 @@ For use and operation, start with the [ODS guide](../source/features/online_data
 This document owns source/Dataset architecture and implementation boundaries; task parameters
 and historical run evidence remain with the demo.
 
+The [ODS data-stream showcase](ods-data-stream-showcase-design.md)
+uses standard SDK/DataLoader video decoding for its offline comparison. That
+example-local reader is separate from the shipped resident numeric backend; its
+visual results cannot be obtained by dropping images or preloading decoded
+datasets through the current numeric loader.
+
 ## Problem and acceptance criterion
 
 `SequenceDataset` supplies one PyTorch context-iteration interface for live and