go2 dds research

dimos/mapping/research/go2/.gitignore

@@ -0,0 +1,36 @@
+# Python-generated files
+__pycache__/
+*.py[oc]
+dist/
+wheels/
+*.egg-info
+
+# Virtual environments
+.venv
+worktrees/
+queue/
+
+# Agent prompt files (generated per-session by launchers)
+CLAUDE.md
+AGENTS.md
+
+# Experimental code/artifacts
+dev/
+
+# Results file (leave untracked)
+results.tsv
+
+# experiment artifacts
+run.log
+
+# large human-debug recordings (kept on disk, not in git; derivable from the mcap)
+human-debug/*.mcap
+human-debug/*.rrd
+
+# point_lio build + per-run outputs (regenerated each run)
+point_lio/build/
+point_lio/Log/*
+!point_lio/Log/.gitkeep
+point_lio/PCD/*
+!point_lio/PCD/do_not_delete_this_file.txt
+point_lio/config/_active.yaml

dimos/mapping/research/go2/autoresearch/lio-1/.gitignore

@@ -0,0 +1,40 @@
+# Python-generated files
+__pycache__/
+*.py[oc]
+dist/
+wheels/
+*.egg-info
+
+# Virtual environments
+.venv
+worktrees/
+queue/
+
+# Agent prompt files (generated per-session by launchers)
+CLAUDE.md
+AGENTS.md
+
+# Experimental code/artifacts
+dev/
+
+# Results / search artifacts (leave untracked)
+results.tsv
+search_log.tsv
+best_config.json
+optuna_study.db
+
+# experiment artifacts
+run.log
+logs/
+
+# large human-debug recordings (kept on disk, not in git; derivable from the mcap)
+human-debug/*.mcap
+human-debug/*.rrd
+
+# point_lio build + per-run outputs (regenerated each run)
+point_lio/build/
+point_lio/Log/*
+!point_lio/Log/.gitkeep
+point_lio/PCD/*
+!point_lio/PCD/do_not_delete_this_file.txt
+point_lio/config/_*.yaml

dimos/mapping/research/go2/autoresearch/lio-1/README.md

@@ -0,0 +1,80 @@
+# lio-autoresearch
+
+An [autoresearch](https://github.com/karpathy/autoresearch)-style autonomous
+experiment loop, adapted from LLM pretraining to **LiDAR-inertial odometry
+tuning**. Give an AI agent a real Point-LIO pipeline and a recorded Go2
+LiDAR+IMU stream, and let it experiment autonomously: modify the config, run
+the LIO, check whether the trajectory agrees better with the robot's onboard
+odometry, keep or discard, repeat.
+
+## How it works
+
+Three files matter (same idiom as the original autoresearch):
+
+- **`evaluate.py`** — fixed data paths, the Point-LIO substrate location, and the
+  evaluation (`evaluate`, the ground-truth metric). Not modified.
+- **`algo.py`** — the single file the agent edits. Holds the Point-LIO `CONFIG`
+  and the run/eval driver. **Edited and iterated on by the agent.**
+- **`program.md`** — baseline instructions for the agent. **Edited by the human.**
+
+The agent does NOT touch the Point-LIO C++ engine (`point_lio/`) — that's the
+fixed substrate, built once by `setup.sh`, analogous to a fixed model framework.
+The agent tunes its configuration.
+
+The metric is **`val_ate_xy`**: the 2D (xy) absolute trajectory error (RMSE,
+meters) of the rigid-aligned LIO trajectory against `robot_odom`, the robot's
+leg-inertial odometry, which loop-closes to ~0.47 m over a 16.5 m path and is
+our rough ground truth. Lower is better. It is **2D only** today; 3D criteria
+(measured step heights from a stairs recording) come later.
+
+## The experiment
+
+The Go2's L1 lidar is sparse (18 lines) and Point-LIO tracks well on straight
+motion but loses lock during in-place rotation (the loop turn-around). The
+research question this harness lets an agent explore: how far can configuration
+tuning alone close the gap to `robot_odom`?
+
+## Quick start
+
+This package runs in the **dimos environment** — it carries no venv of its own.
+The C++ build deps (cmake, eigen, pcl, yaml-cpp, boost) come from the Nix dev
+shell defined in `../flake.nix`; the python deps (numpy, matplotlib) and
+`dimos.get_data` come from the dimos `.venv`. The LIO input + ground truth are
+pulled from the dimos LFS data store (`go2dds_data1`) by `get_data` on first run.
+
+```bash
+# 1. Enter the dev shell (C++ build deps layered on the dimos dev shell)
+cd ..            # -> dimos/mapping/research/go2/autoresearch
+nix develop
+
+# 2. Build the substrate + sanity check (pulls the data via get_data)
+cd lio-1
+./setup.sh
+
+# 3. Run a single baseline experiment (~1-3 min)
+python algo.py
+```
+
+If those work, you're ready for autonomous mode: point your agent at
+`program.md`.
+
+## Project structure
+
+```
+evaluate.py        — data paths (via get_data), substrate location, evaluation (do not modify)
+algo.py          — Point-LIO CONFIG + run/eval driver (agent modifies this)
+program.md        — agent instructions
+setup.sh          — environment prep (build point_lio; checks the dimos venv)
+config/v2_imu.yaml— the baseline config, for reference
+point_lio/        — the fixed Point-LIO C++ engine (built by setup.sh)
+human-debug/      — convert script, raw .mcap, .rrd rerun — HUMAN USE ONLY, not in the loop
+```
+
+Data lives in the dimos LFS store, not here: `data/go2dds_data1/` holds
+`20260529-185959.mcap` (LIO input — Point-LIO reads it directly) +
+`gt_robot_odom.tsv` (ground truth), fetched via `get_data("go2dds_data1/...")`.
+
+## License
+
+The `point_lio/` engine is GPL (see `point_lio/LICENSE`); it derives from
+unitreerobotics/point_lio_unilidar. The harness files are MIT.