RNCOA Aggregated Collision Loss
Implementation & benchmark of arXiv:2607.02797 — RCOA Extension and Applications
Source paper: RCOA Extension and Applications arXiv:2607.02797 ⓘ · analyzed Aug 30, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Attach an RNCOA-inspired collision loss to a neural trajectory or control-policy head that predicts the pose of a rigid vehicle over time. For each obstacle and time step, aggregate the signed obstacle coordinates of all body vertices using max/min operators, and introduce two nonnegative side slacks whose sum is constrained to at most one. This models the disjunctive fact that the complete body should lie on one admissible side of an obstacle rather than independently penalizing every vertex.
Formulas
Mathematical statement
The paper's RNCOA constraints are $-\max_{i\in\mathcal V}\{j_i\}\le -j^o_{\min}+M_1\gamma_1$ and $\min_{i\in\mathcal V}\{j_i\}\le j^o_{\max}+M_2\gamma_2$, together with $\gamma_1+\gamma_2\le1$. Here $\mathcal V$ is the set of vehicle vertices, $j_i$ is the scalar signed obstacle-coordinate value for vertex $i$, $j^o_{\min}$ and $j^o_{\max}$ are obstacle interval boundaries, $M_1,M_2>0$ are big-M relaxation constants, and $\gamma_1,\gamma_2\ge0$ are side-specific slack variables. The max/min aggregation couples all vertices, while the sum constraint prevents relaxing both sides simultaneously. For differentiable neural training, use $\operatorname{smax}_\tau(j)=\tau\log\sum_i\exp(j_i/\tau)$ and $\operatorname{smin}_\tau(j)=-\tau\log\sum_i\exp(-j_i/\tau)$, where $\tau>0$ controls smoothing. Penalize positive violations and slack usage. The paper's objective also motivates adding a task-tracking term such as $\sum_t[w_1\|p_y^t-r_y\|_1+w_2\|p_z^t-r_z\|_1+w_3(\gamma_1^t+\gamma_2^t)]$, where $p_t$ is predicted position, $r_t$ is reference position, and $w_1,w_2,w_3$ are weights.
Implementation notes
Integrate this at the loss layer of a neural trajectory predictor, model-predictive policy, or RL actor. Let the network output poses $(p_t,R_t)$ for $t=1,\ldots,T$ and two slack logits per obstacle and time step; map the logits to nonnegative $\gamma_{1,t,o},\gamma_{2,t,o}$ with softplus or sigmoid activations. Precompute the vehicle body-frame vertices $v_i$. Transform them into world coordinates using $q_{t,i}=p_t+R_t v_i$. For each obstacle $o$, compute a scalar signed coordinate for every transformed vertex. For an axis-aligned obstacle, use the relevant coordinate directly; for a separating-plane representation, use $j_{t,o,i}=n_o^\top q_{t,i}$, where $n_o$ is the obstacle normal and $j^o_{\min},j^o_{\max}$ are the obstacle bounds along that normal. Compute the soft maximum and soft minimum, then evaluate the two ReLU violations in the displayed $L_{\mathrm{RNCOA}}$ formula. Add the loss to imitation error, task reward loss, or the actor objective. Start with $\tau=0.1$ and anneal to $0.01$; use exact max/min and mesh intersection tests for validation. The paper's math determines the coupled extrema and exclusive-slack structure. Estimate only geometric bounds and tune $M_1,M_2,\lambda_\gamma,\lambda_c$ empirically; choose each $M$ larger than the maximum normalized coordinate violation. First test a small MLP trajectory policy on 2-D rectangle navigation with a four-vertex rectangular body. Compare against independent per-vertex hinge penalties at equal FLOPs. Record exact collision rate, minimum clearance, gradient variance, and training throughput. Success means lower collision rate at equal tracking error, fewer cases where one body vertex collides despite other vertices being safe, and modest computational overhead.
Verification
Stage 1 · Toy mechanism gate: Failed ✗
Stage 2 · Mechanism transferred to benchmark: Not tested
Stage 2 · Practical benchmark result: Not run
Methodology: Toy-system gate first; the benchmark stage runs only after a pass. How verification works
Stage 1 — Mechanism check agent confidence 10/10
Built rncoa_experiment.py implementing the stated RNCOA soft-extrema loss, independent per-vertex hinge baseline, exact 1D body collision test, smoothing verification, random confusion test, and direct optimization comparison. The math check showed soft max/min errors falling from about 1.34e-3 at tau=0.2 to numerical zero at tau<=0.05, but 100% of sampled colliding bodies had zero RNCOA loss while safe bodies were often penalized. In optimization, RNCOA remained at the colliding initialization whereas the baseline reduced its hinge loss, although its independent objective still permits a straddling body; therefore the idea did not work with the authoritative signs.
- Agent confidence
- 10/10
- Baseline
- Initial independent hinge 1.60 -> final 0.0; optimized vertices [-0.25335, 1.27876, 1.27606, -0.24983], exact collision=true because independent vertex penalties do not enforce whole-body separation.
- Idea
- Initial/final stated RNCOA loss 0.0; vertices stayed [0.45, 0.55, 0.65, 0.35], exact collision=true. Random test: zero-loss rate on collisions 1.0 (5219/5219), while 1.05% of safe samples had zero loss.
Limitations: No full neural trajectory MLP, 2D rotated rigid-body experiment, throughput, gradient-variance, or GPU benchmark was run because the authoritative loss failed the prerequisite geometric sanity check. The test uses a one-dimensional obstacle coordinate and zero learned slacks; correcting the inequality directions would require a separate experiment and is not treated as validation of the supplied formula.
How to run: python3 rncoa_experiment.py
Verdict computed by deterministic test code from paired-seed statistics — not by the language model.
Artifacts
- report.md 1.8 KB View
- results.json 2.5 KB View Raw
- rncoa_experiment.py 5.6 KB View Raw