Nav pt3: Enable PGO topics, Optimize PGO, and add KITTI-360 benchmark scaffolding#2099
Open
jeff-hykin wants to merge 92 commits into
Open
Nav pt3: Enable PGO topics, Optimize PGO, and add KITTI-360 benchmark scaffolding#2099jeff-hykin wants to merge 92 commits into
jeff-hykin wants to merge 92 commits into
Conversation
… rrb
Native module (cpp/main.cpp) now publishes two new streams on every
keyframe: GraphNodes3D for keyframe optimized poses, LineSegments3D for
odometry (traversability=1.0) and loop-closure (0.4) edges. Both wire
through SimplePGO::keyPoses() + historyPairs() — no changes needed to
simple_pgo.{h,cpp} since the accessors already exist. Native binary
rebuilt cleanly via nix build .#default --no-write-lock-file.
Python (pgo.py) declares matching pgo_graph_nodes / pgo_graph_edges Out
streams so the rerun bridge auto-discovers and logs them.
nav_stack_rerun_config() now picks _agentic_debug_rerun_blueprint when
agentic_debug=True — an rrb.Horizontal layout with a 3D pane and a
dedicated top-down pane (both Spatial3DView over origin="world", named
"3D" and "top_down" so dimos-viewer persists camera state separately).
demo_better_pgo_viz.py composes the cross-wall sim blueprint with
agentic_debug=True so the new layout + pose graph render together. Used
for manual screenshot validation.
Adds visual_override entries for world/pgo_graph_nodes and world/pgo_graph_edges that mirror the existing FAR pattern: when agentic_debug=True, the PGO pose graph renders at z=_AGENTIC_DEBUG_LIFT (3.0m) instead of the default 1.7m, with slightly larger node radii (0.15) and edge thickness (0.06) so the green keyframe trajectory stands out clearly above the terrain cloud in the top-down pane. Verified visually via demo_better_pgo_viz with the cross-wall sim — green keyframe nodes + edges are now plainly identifiable above terrain in both the 3D and top_down rerun panels.
rerun's Spatial3DView doesn't have a top-down camera API, so the "top_down" pane introduced in a7a9be9 was just a duplicate 3D view. Drop _agentic_debug_rerun_blueprint and use _default_rerun_blueprint unconditionally — the agentic_debug lift on visual_override is what actually makes the pose graph and nav markers readable from any angle.
C++ side (main.cpp): when searchForLoopPairs sets m_cache_pairs (i.e. this keyframe will be incorporated into iSAM2 with a loop factor), snapshot the current global poses before smoothAndUpdate. After the update, build a nav_msgs::Path-encoded LoopClosureDeltas message: position = post.t - r_delta * pre.t, orientation = quaternion(post.R * pre.R^T). Publish on the new pgo_loop_closure topic. Stderr logs the event count for live observability. Python side (pgo.py): declare pgo_loop_closure: Out[NavPath] so the new topic is registered alongside corrected_odometry/pgo_tf/etc. Slow test (test_pgo_loop_closure.py): replays og_nav_60s through the native binary with permissive thresholds (loop_time_thresh=5s, min_loop_detect_duration=1s, loop_search_radius=2m, loop_score_thresh=0.5) so the recording reliably triggers loop closures. Subscribes to pgo_loop_closure, logs each event the moment it arrives (event #, poses_length, frame_id, first delta), and after the run validates each event has >0 poses, finite translations (<100m), and unit-norm quaternions (drift <0.05). Stdout from a run shows 19 events, sizes 10..35, max |t|=0.0013m, max |q|-1|=1e-6 — exactly the small-nudge profile expected from a self-consistent recording.
Replaces the kdtree-on-keyframe-positions loop search with a Scan
Context (Kim & Kim 2018) descriptor-based pipeline:
1. addKeyPose now also caches a polar-binned (20 rings × 60 sectors)
max-z descriptor + the per-row mean "ring key" for each keyframe.
The descriptor is appearance-based and pose-independent, so it
keeps working even when odometry has drifted enough that the new
keyframe is no longer "near" its old neighbours in pose-space.
2. searchForLoopPairs first asks Scan Context for a candidate:
ring-key L2 distance ranks all past keyframes, top-K are scored
by column-shifted cosine distance on the full descriptor, the
best below the threshold (default 0.4) is the candidate. The
winning column shift is also converted to a yaw rotation and used
to seed ICP, which dramatically improves convergence on revisits
that arrive at a different heading from the original pass.
3. Position-based search is retained as a fallback when SC is
disabled or finds nothing, so existing behaviour is preserved.
Replaces ~50 lines of position-search with ~30 lines of SC retrieval
in searchForLoopPairs; new scan_context.{h,cpp} (~150 lines, MIT
attribution to upstream irapkaist/scancontext concepts but no source
copied) implements the descriptor + distance.
Side-effect: this makes on-start relocalization a small follow-up
addition — descriptors + ring-keys + poses are now per-keyframe state
that can be serialised, and the SC search path already does
"appearance-based pose recovery without an initial pose guess."
Verified via test_pgo_loop_closure.py: 17 loop-closure events fired
across the og_nav_60s rosbag (was 19 with naive position search; SC
is more selective and rejects two borderline-position matches that
weren't actually visual revisits). All events have valid shape + tiny
quaternion/translation deltas as expected for a self-consistent bag.
…n search misses Adds CLI args to expose Scan Context config on the native binary (--use_scan_context, --sc_n_rings, --sc_n_sectors, --sc_max_range_m, --sc_top_k, --sc_match_threshold). New slow test test_pgo_synthetic_drift.py: - Synthesises a 4-wall point-cloud room with two distinctive interior columns (so the scene isn't rotationally symmetric). - Generates an out-and-back trajectory: drives east 8m then returns to the origin, heading unchanged. - Injects DRIFT_AT_REVISIT_M = 5m of additive y-drift into the reported odometry, ramped linearly with travelled distance. The body-frame scan stays byte-identical between first and second visit (same true sensor view of the same scene); the odom pose at revisit is 5m offset. - Runs the native PGO binary twice over the same input: * use_scan_context=true → expect ≥1 loop event * use_scan_context=false → expect 0 loop events (drift >> 1m radius) - Dumps PGO stderr after each run for diagnostics. Result: SC fires 10 loop closure events on the synthetic trajectory; position-based search fires 0 — exactly the demonstration of why we swapped to appearance-based place recognition. Both assertions pass. Verifies the core SC value prop: appearance-based place recognition doesn't depend on the (drifted) pose, so it keeps working when the odometry has wandered far enough that the kdtree-on-positions search no longer finds neighbours.
Test files now use setup_logger() / logger.info(...) per the fix_nits rule "no print() calls in tests; use logging if diagnostics are genuinely needed." Matches the existing test_pgo_rosbag.py convention. Also drops the now-unused sys import. Also clears a stale docstring on demo_better_pgo_viz.py: it claimed the demo enabled a "horizontal 3D + top-down panes" layout, which was reverted in 1801759 — rerun's Spatial3DView didn't support an initial camera angle (rrb.EyeControls3D existed at the time but wasn't used). The remaining value of agentic_debug=True is the visual override lift, which the new docstring describes accurately. No behavioural change. Tests still pass.
Sweep over names introduced by the better_pgo work that hit fix_nits
"expand mod -> module" rule:
- scan_context: cfg -> config (param + 12 call-sites); d (return val) ->
descriptor in make_descriptor/make_ring_key/make_sector_key; pt -> point
in the descriptor build loop; zf -> point_z (float cast); q_col/c_col
-> query_column/candidate_column; q_norm/c_norm -> query_norm/
candidate_norm; cj -> shifted_j; d (in best_distance return loop) ->
distance with min_distance for the running best.
- simple_pgo: desc -> descriptor on the per-keyframe cache; k ->
top_k_count for the partial-sort bound; structured-binding `auto [d,
shift]` -> `auto [distance, shift]`.
- main.cpp: kp -> keyframe; ps -> pose_stamped (build_graph_nodes and
build_loop_closure_deltas); a/b -> start/end and p1/p2 ->
start_pose/end_pose in append_segment; n -> count for the loop bound;
lc_msg -> loop_closure_msg at the publish site.
- tests: ps -> pose in the validate loop (test_pgo_loop_closure);
c,s -> cos_yaw,sin_yaw in _yaw_rotation (test_pgo_synthetic_drift).
Names that intentionally stay short are the math-convention ones:
r/t for SE(3) rotation+translation, q for quaternion, i/j as loop
indices, idx as keyframe index, ts as timestamp, dt for time delta,
tx/ty/tz/qx/qy/qz/qw for component decomposition. The fix_nits rule
calls out mod/lc as the target pattern; expanding the math-notation
names would make the code less readable, not more.
Also drops one section-label comment ("# Log each event the moment it
arrives.") whose adjacent function name already conveys the same and
one in-loop "# node_type 1 = odom/robot" that repeats info already
stated in the function-level docstring.
Native binary rebuilt + slow test still passes (17 events, all valid).
Drops in the wiring for evaluating the PGO native module on KITTI-360. Cannot run end-to-end yet — the dataset is gated behind a registered login at cvlibs.net so the data download is a manual user step. What's here: - kitti360_loader.py: parses the KITTI-360 directory layout (data_3d_raw + data_poses + calibration); composes per-frame lidar→world pose by chaining cam0_to_world ⊕ inv(velo_to_cam). Exposes a frame iterator + scan_xyz(frame_id). - loop_groundtruth.py: LCDNet/KITTI-convention groundtruth (≥50 frame gap, ≤4m radius), order-agnostic scoring of detected pairs. - run_kitti360_benchmark.py: argparse CLI, spawns the native binary on private LCM topics, plays (registered_scan, odometry) from disk, subscribes to pgo_graph_edges to extract loop-closure pairs (via traversability ≈ 0.4 segments) and pgo_loop_closure for delta event counts. Writes JSON. - README.md: download instructions for the official "Test SLAM 3D" 12 GB package, published SOTA reference numbers from LCDNet + ISC papers (LCDNet 0.91-0.93 AP, Scan Context 0.62-0.78 AP), expected ballpark for our minimal SC port.
Codecov Report❌ Patch coverage is 📢 Thoughts on this report? Let us know! |
Contributor
Greptile SummaryThis PR enables PGO topics (
Confidence Score: 4/5Mostly safe to merge; benchmark and message-type plumbing are well-structured, but simple_pgo.cpp still exits early when SC finds a candidate and ICP rejects it, skipping the position-based fallback for that keyframe. Most issues flagged in earlier rounds have been addressed: SC params are now in PGOConfig, metrics are query-level, timestamps raise on mismatch, try/finally guards are in place. Two items from prior reviews remain open: simple_pgo.cpp returns without attempting the position fallback when ICP rejects an SC candidate, and SyntheticDriftPlaybackModule stores _playback_error but never exposes it via @rpc so _run_pgo cannot detect PGO startup failures. dimos/navigation/nav_stack/modules/pgo/cpp/simple_pgo.cpp (searchForLoopPairs early-return after ICP failure) and dimos/navigation/nav_stack/modules/pgo/test_pgo_synthetic_drift.py (SyntheticDriftPlaybackModule missing playback_error @rpc). Important Files Changed
Sequence DiagramsequenceDiagram
participant PB as Kitti360PlaybackModule
participant PGO as PGO (C++ iSAM2+SC)
participant SC as PoseGraphScoringModule
PB->>PGO: odometry (send_timestamp, pose)
PB->>PGO: registered_scan (world-frame cloud)
Note over PGO: unregister_input: world to body
Note over PGO: addKeyPose to iSAM2 factor
Note over PGO: searchForLoopPairs (SC + ICP)
alt loop found
Note over PGO: snapshot pre_poses
Note over PGO: smoothAndUpdate (iSAM2)
PGO->>SC: loop_closure_event (GraphDelta3D)
end
PGO->>SC: pose_graph (Graph3D, every keyframe)
PGO->>PB: corrected_odometry (signals PGO alive)
Note over SC: _on_pose_graph: map node.id to pose.ts
Note over SC: _timestamp_to_frame (1ms slop)
Note over SC: _score_pairs: query-level TP/FP/FN
Note over PB: is_finished() returns True
PB-->>SC: get_results() returns precision/recall/F1
Reviews (43): Last reviewed commit: "revert" | Re-trigger Greptile |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
- run_kitti360_benchmark: type the scipy Rotation.as_quat result to silence no-any-return. - demo_better_pgo_viz: annotate build_blueprint() -> Blueprint.
shift comes from best_distance, which scans [0, n_sectors-1], so the raw yaw is in (-2pi, 0] and `yaw > M_PI` can never fire. Only the negative-wrap guard is needed to normalize into [-pi, pi].
dimos/ disallows __init__.py files (test_no_init_files) — the empty one in pgo/benchmark/ slipped in with c7fd631 and was tripping the 3.14 test job.
mypy can't infer parameter types on a lambda subscribed to LCM; lift the body into a tiny factory function with explicit Callable[[str, bytes], None] signature so the lint job passes.
sklearn doesn't ship a py.typed marker; the new place_recognition_ap benchmark is the only sklearn user in dimos, so a per-import ignore is the smallest fix to unstick the lint job.
Naming sweep across dimos/navigation/nav_stack/modules/pgo per fix_nits review conventions: expanded short locals (frac, true_pos, dt, ts, pos, yaw, msg, sub, idx, fid, gt, cfg, desc, dots, sims, dists, pa/pb, …) to full descriptive names in the benchmark scripts, the synthetic-drift test, and the loop-closure test. Greptile P1 fixes on PR #2099: - c2: benchmark sender and the timestamp_ms→frame_id cache now share a single _compute_send_timestamps source of truth. Previously the cache was keyed by raw KITTI timestamps while the runner sent max(raw_ts, 1.0 + index*0.001), so early-frame endpoints in PGO loop edges never matched the cache and were silently dropped — deflating recall without any warning. - c3: load_kitti360_sequence now raises on .bin/timestamps.txt length mismatch instead of silently leaving timestamps={}, which previously caused the benchmark to report recall=0 with no indication that the timestamp file was unusable. - c1 follow-up: rename the misspelled C++ field loop_score_tresh → loop_score_thresh in simple_pgo.{h,cpp} and main.cpp. The CLI flag was always spelled correctly; this is cosmetic but removes the source of confusion greptile (correctly) flagged. New regression test (test_pgo_synthetic_drift.py): test_scan_context_catches_reverse_loop drives the synthetic robot out 8m facing east, turns 180°, and drives back facing west. Body-frame scans on the return leg are rotated 180° relative to outbound, so this exercises the init_guess fix in searchForLoopPairs (yaw rotated about the source keyframe instead of the world origin). Reverting that fix reproduces the failure. New runners: - benchmark/place_recognition_ap.py: apples-to-apples place-recognition AP eval against published Scan Context numbers. Reports AP=0.97-0.98 on seq 02/04/08 of the Test SLAM split, with precision 1.000 at the Kim & Kim 0.13 threshold. - benchmark/smoke_test.py: ~10s liveness probe that subscribes to all six PGO output topics, captures stderr, and prints per-topic message counts plus a one-line verdict — used to distinguish "PGO crashed" vs "no keyframes" vs "no loops" vs "alive" during debugging. The benchmark runner also now captures PGO's stderr and dumps it behind --print-stderr; previously its diagnostic prints (keyframes, loop-closure events) were discarded.
jeff-hykin
changed the title
…raph3D] Bumps the upstream pin to dimos-module-far-planner v0.7.0, which collapses the two-stream (GraphNodes3D + LineSegments3D) visibility-graph output into a single graph: Out[Graph3D] stream. Node metadata_id matches the previous GraphNodes3D node-type enum (0=normal, 1=odom, 2=goal, 3=frontier, 4=navpoint). Edge metadata_id encodes traversability tier (0=non-traversable, 1=partial, 2=both endpoints traversable) instead of the float orientation.w convention. nav_boundary stays on LineSegments3D — it's a polyline (collision boundaries with no graph topology), not a graph. Validated end-to-end with test_far_planner_rosbag.py against v0.7.0.
…p GraphNodes3D Collapse PGO's two-stream nav_msgs/Path encoding into a single pose_graph: Out[Graph3D] snapshot per keyframe cycle. Each node carries a full SE(3) PoseStamped (previously rotation was zeroed), a stable uint64 id (= keyframe index), and metadata_id = NODE_KEYFRAME. Edges reference nodes by id and carry metadata_id = EDGE_ODOMETRY (0) or EDGE_LOOP_CLOSURE (1) — replaces the orientation.w == 0.4 traversability tag previously used to mark loop closures. Touches: * pgo.py / specs.py: Out[NavPath] x 2 → Out[Graph3D] x 1. * cpp/main.cpp: build_graph_nodes + build_graph_edges + append_segment helpers collapse into build_pose_graph() returning dimos::Graph3D. * cpp/msgs/GraphNodes3D.hpp deleted (vendored copy no longer needed). * scoring.py: filter loop edges via metadata_id == EDGE_LOOP_CLOSURE instead of orientation.w ≈ 0.4. Endpoint frame_id lookup goes through node_id → node.pose.ts → timestamp_to_frame, removing the loop_closure_traversability / traversability_tolerance config. * nav_stack/main.py: _pose_graph_nodes_colors_debug + _pose_graph_edges_colors_debug → _pose_graph_colors_debug, calling Graph3D.to_rerun_multi to render nodes + edges under separate rerun sub-paths. * runner.py / README.md docstring updates. Deletes dimos/msgs/nav_msgs/GraphNodes3D.py — no longer has consumers after this commit and the matching far_planner switch. Validated by: * test_Graph3D.py round-trip + wire-layout pin tests pass. * test_far_planner_rosbag.py (uses the same Graph3D Python type for far_planner's graph stream) passes. * benchmark_kitti360_smoke deterministically produces keyframes across 5 consecutive runs at publish_interval_sec=0.02 (50 Hz), including publish_interval_sec=0.0 (unbounded).
The smoke test isn't a demo (it runs PGO end-to-end with KITTI-360 playback and reports per-topic message counts as a liveness check), and it's specific to KITTI-360 — the new name reflects both. Also matches the sibling benchmark module names under ``dimos/navigation/nav_stack/benchmarks/pose_graph_kitti360/``. Pure rename + topic count updates: the TopicCounterModule subscribes to pose_graph (Graph3D) instead of pose_graph_nodes + pose_graph_edges (matching the PGO output collapse in the previous commit), and the verdict messages collapse accordingly.
Earlier rename to test_pgo_loop_correction_delta.py (29dfafe) was out-of-sync with main, where the file went the opposite direction. Restore the upstream filename to avoid a manual rename-direction conflict on the next merge. Internal symbols (LoopCorrectionDeltaRecorderModule, the loop_correction_delta stream attribute, etc.) stay — those describe what the stream actually carries (per-keyframe SE(3) deltas, not loop closures themselves).
… into jeff/feat/better_pgo
`topic-counter-module` path updated to reflect the demo_benchmark_kitti_smoke → benchmark_kitti360_smoke rename (commit f97e1bd).
New nav_msgs.GraphDelta3D: two aligned arrays — a list of nodes (reusing Graph3D.Node3D byte-for-byte) and a list of SE(3) Transforms, one per node. ``transforms[i]`` is the delta about to be applied to ``nodes[i].pose``: ``post = transforms[i] * nodes[i].pose`` (left-multiply). Layout (big-endian, mirrors Graph3D conventions): u8 node_count f8 timestamp Node3D nodes[node_count] // same wire bytes as Graph3D nodes Transform transforms[node_count] // 7×f8 (translation xyz + quat xyzw) Test pin (test_node_layout_matches_graph3d) asserts that a GraphDelta3D.nodes[i] is bit-identical to a Graph3D.nodes[i] for the same Node3D, so consumers can correlate ids across the two messages. PGO wiring: * pgo.py / specs.py: ``loop_correction_delta: Out[NavPath]`` → ``loop_closure_event: Out[GraphDelta3D]``. Renamed too — the stream carries a graph-mutation event, not just a delta payload. * cpp/main.cpp: build_loop_correction_delta (returned nav_msgs::Path with position-encodes-delta hack) → build_loop_closure_event returning dimos::GraphDelta3D. Each (node, transform) pair carries the pre-smooth keyframe (id, pose, ts) and the iSAM2-applied delta. * cpp/msgs/GraphDelta3D.hpp vendored. Downstream consumers updated to subscribe ``loop_closure_event: In[GraphDelta3D]``: * PoseGraphScoringModule * test_pgo_loop_closure recorder (full payload validation rewrite — unit-quaternion + finite-translation checks now read each transform.rotation / transform.translation instead of pose.orientation / pose.position) * test_pgo_synthetic_drift counter * benchmark_kitti360_smoke TopicCounter * runner.py docstring 11/11 unit tests pass (5 GraphDelta3D layout, 5 Graph3D layout, 1 blueprints registry). PGO native rebuilt — new topic ``loop_closure_event`` confirmed in binary symbols.
jeff-hykin
commented
May 17, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #2007
Changes
How to Test
without benchmark
With benchmark
The benchmark is KITTI 360 - a standard dataset but its stuck behind a registration page (no public download). Its also 12Gb. https://www.cvlibs.net/datasets/kitti-360/user_login.php
There's no good 100% public loop closure benchmark that I could find.
Not in this PR
Contributor License Agreement