This log records design issues and key decisions for the project. Dates refer to the review in which the issue was first raised and the review in which its resolution was reported.
| Item | Date initiated | Date resolved | Participants | Description | Options | Resolution | Justification |
|---|---|---|---|---|---|---|---|
| 1 | 2026-09-09 (PR7) | 2026-09-23 (PR8) | TBD | Choose the onboard VLM for button and panel grounding. | Cosmos-Reason 8B; Qwen3-VL-Instruct 4B. | Cosmos-Reason2-8B selected; Qwen3-VL 4B kept as fallback. | Highest accuracy on the three panel/button queries (17/18 vs. 16/18). |
| 2 | 2026-09-23 (PR8) | 2026-09-23 (PR8) | TBD | Choose a deployment path for the onboard VLM within latency limits. | vLLM BF16 + FP8 KV cache; vLLM FP8; TensorRT Edge-LLM FP16, FP8, NVFP4 + FP8 vision, NVFP4 + FP16 vision. | TensorRT Edge-LLM FP8 with a resident engine and a tight output schema. | 19/19 parsed responses at 2.08 s mean latency; vLLM FP8 produced corrupted output and NVFP4 lost accuracy (11–12/19). End-to-end grounding latency dropped from 7.18 s to ~1.67 s. |
| 3 | 2026-04-27 (CDR) | 2026-10-21 (SDR) | TBD | Perception depends on Wi-Fi to external compute; network drops interrupt the mission. | Keep external compute; run perception onboard. | Perception (VLM grounding, segmentation, 3D tracking) moved onboard to the Orin/Thor with TensorRT. | Removes the network dependency; 100% target grounding (51/51) at ~12 FPS tracking. |
| 4 | 2026-09-09 (PR7) | 2026-10-21 (SDR) | TBD | Camera-to-robot transform error is too large for manipulation. | Camera–LiDAR calibration only; online calibration / system identification; hand-eye calibration. | New camera mount and camera–LiDAR calibration, followed by hand-eye calibration (ICP of the RGB-D hand cloud against the FK hand mesh). | The LiDAR pose in the Unitree URDF is wrong, so relative calibration alone left a residual; hand-eye calibration corrects the camera directly. |
| 5 | 2026-04-27 (CDR) | 2026-10-21 (SDR) | TBD | Unstable button detection prevents reliable pressing. | Per-frame VLM detection; onboard object tracking; pose filtering. | Onboard object tracking with position holding; later an EKF on body pose and a KF on target pose (PR10). | Target pose is much more stable during the approach. |
| 6 | 2026-04-27 (CDR) | 2026-10-21 (SDR) | TBD | Poor base servoing harms button pressing. | Tune the existing holonomic controller; change motion model. | Calibrated the IMU, removed lateral velocity and switched to a non-holonomic planner. | Reduces lateral drift and keeps the panel in a reachable pose. |
| 7 | 2026-09-09 (PR7) | 2026-10-21 (SDR) | TBD | Arm motion for button pressing is not repeatable and planning is slow. | Further motion tuning; slower approach with a pause; planner optimization. | Path smoothing and speed tuning, a slower final approach with a one-second pre-grasp pause, and an optimized planner. | Planning time 2194–2583 ms → 682–838 ms; IK calls 22 → 11; goal residual 186 µm → 39 µm (offline). |
| 8 | 2026-10-21 (SDR) | 2026-10-27 (PR10) | TBD | Arise-SLAM cannot map across floors because of IMU drift in the elevator. | Arise-SLAM; Elevator-LIO. | Replaced Arise-SLAM with Elevator-LIO for state estimation. | Elevator-LIO maps across floors. |
| 9 | 2026-10-21 (SDR) | 2026-10-27 (PR10) | TBD | The robot must relocalize in a saved map before navigating. | — | Added GICP-based relocalization. | Relocalized successfully with a structure score of 0.97. |
| 10 | 2026-10-21 (SDR) | 2026-10-21 (SDR) | TBD | Project scope after losing one team member. | Keep semantic mapping; remove it from scope. | Semantic mapping removed from scope. | Protects the elevator mission with the remaining team capacity. |
| 11 | 2026-10-07 (PR9) | Pending | TBD | How to chain skills into a full elevator mission. | Hand-written rule-based chain; LLM-based agentic chaining. | Hand-written baseline runs on the real robot; agentic chaining completed its first end-to-end rollout in simulation but is not yet tested on hardware. | Keep a demonstrable baseline while the agentic approach matures; the agent depends on the network and once proposed entering with the door only 69% open. |
| 12 | 2026-10-07 (PR9) | Pending | TBD | Hardware breakdowns (Dex3 hands, robot downtime) reduce test time. | Repair parts; provision a new robot; use Dex1 hands. | More than ten Dex3 hands repaired and the robot recovered; provisioning an additional robot is in discussion. | Restores test time; simulation testing used during downtime. |
| 13 | 2026-10-27 (PR10) | Pending | TBD | The robot still occasionally misses the button. | — | Pending: collect success rates and separate perception error from control error. | Root cause not yet identified. |
| 14 | 2026-10-27 (PR10) | Pending | TBD | Local planner produces poor locomotion plans. | Find an alternative planner; tune collision-avoidance parameters. | Pending. | — |
| 15 | 2026-10-27 (PR10) | Pending | TBD | Elevator-LIO may not detect the moment of entering the elevator. | Add another gate to the IMU-correction activation condition. | Pending. | — |